smolmaxxing
maxxing the ability to run different LLMs on the smallest devices possible
4 repos · 49 papers · 43 scholar
Showing 96 of 96 items
maxxing the ability to run different LLMs on the smallest devices possible
Sushi engine runs very high quality quant model in Apple Silicon, using both STQ and EXL3 format with special tuning technology to achieve frontier sub 4 bits compression.
Live person detection with a Prophesee GenX320 event camera on an STM32F746 (Cortex-M7): firmware, recording, curation, simulation, training and evaluation tools of the N-People dataset
Memory decides which LLMs an edge device runs, and how fast: capacity on a $4$ GB Jetson Nano, bandwidth on an $8$ GB Jetson Orin Nano Super. Deployed formats (GPTQ, AWQ, $\texttt{llama.cpp}$'s K- and I-quants) spend equal bits on every integer code of a matrix, although the codes are far from uniform. $\texttt{LazyWeights}$ keeps these formats but tunes GPTQ so the codes compress, Huffman-codes them losslessly in interleaved lanes, and decodes them on demand in a small unit between DRAM and L2, so GPU kernels run unchanged and DRAM moves only coded bytes. At $1\%$ and $3.5\%$ perplexity budgets, coded Phi-3-mini, Mistral-7B and SmolLM-1.7B are $5.7-31.4\%$ smaller than the best $\texttt{llama.cpp}$ format. A bit-exact $0.21$ mm$^2$ ASIC decoder unit ($1$ GHz, $7$ nm) cuts Orin latency and energy per token by up to $36\%$ and $38\%$ ($54\%$ and $44\%$ with prompt-lookup speculation). On the Nano, $\texttt{LazyWeights}$ needs $58-64\%$ less time and $53-57\%$ less energy per token than CPU $\texttt{llama.cpp}$, and runs a streamed Mistral-7B 6.6$\times$ faster at $3.89$ bits per weight. A Huffman decoder beats an rANS decoder in rate and area at equal throughput; the ASIC decoder unit runs bit-exact on a Zynq UltraScale+ FPGA (AMD ZCU102).
Frequency-domain multiplexing is a fundamental technology for readout of cryogenic detector arrays on a single transmission line. Synthesizing a large number of static carrier tones for resonator stimulation eventually drives the cryogenic low-noise amplifier beyond its P1dB, introducing detector signal distortion. Scaling array sizes to O(10^6) detector pixels therefore demands minimizing the total power on the transmission line by continuous tracking of the resonances, which requires the readout electronics to be capable of updating the carrier tones dynamically. This work presents an scalable FPGA-based implementation for carrier signal synthesis based on a numerically controlled oscillator and a polyphase filter bank, enabling real-time frequency adaptivity as well as frequency resolution of 4 mHz to precisely tune the operating point, while simultaneously reducing resource usage compared to the current memory-based implementation.
Symbolic reasoning at scale requires representations that are both expressive and hardware-efficient. Hyperdimensional Computing (HDC) provides a promising brain-inspired framework, encoding concepts as high-dimensional vectors composed through efficient algebraic operations. However, the widely used Fourier Holographic Reduced Representation (FHRR) relies on commutative binding, which cannot encode order or directionality. Prior FPGA/ASIC accelerators recover this through costly permutation logic, incurring irregular memory access and high control overhead. Generalized Holographic Reduced Representations (GHRR) instead redefine binding as non-commutative matrix multiplication, naturally encoding sequences, graphs, and hierarchies without permutations-but at the cost of turning every element into a matrix, introducing compute and storage overhead that limits scalability. In this work, we propose a flattening methodology that preserves GHRR's matrix-based encoding while executing training and inference directly in vector space, functionally equivalent to FHRR inference yet free of permutation logic. We further design a custom 28 nm ASIC that fuses binding and similarity into a unified complex-valued datapath, with dual-DMA streaming and runtime normalization for accurate inference. Against a PyTorch baseline on an NVIDIA RTX 4090 GPU, the prototype delivers 1.36x-1.56x higher throughput and 16.2x-18.6x better energy efficiency.
Fault-tolerant quantum computing requires classical decoders that can keep up with syndrome data streams and supply correction information and logical measurement outcomes to the quantum control system. Meeting these requirements at scale requires high throughput while controlling the end-to-end latency from readout to usable correction. In this paper, we organize decoding algorithms, execution strategies, and code constructions using common classifications and examine their performance and trade-offs under the reported evaluation conditions. We distinguish decoder-level hierarchy, where multiple decoding stages or resources cooperate to decode a single code, from code-level hierarchy, where the code construction itself introduces multiple decoding levels. We first review single-code hierarchical decoders and their cascaded predecoding and windowed, streaming, and distributed implementations. We then study hierarchical and concatenated codes, focusing on how decoding is coordinated across code levels through hard decisions, soft reliability information, or syndrome information. We also examine how qubit modalities shape decoding workloads and timing requirements, and review how different decoding platforms meet these requirements, including central processing units (CPUs), graphics processing units (GPUs), field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), and cryogenic platforms. By assessing design trade-offs, this paper clarifies connections across the literature and provides a shared foundation for QEC design exploration. It also identifies gaps in performance reporting and system integration to guide future research toward practical quantum error correction.
Protocol bridge verification is a persistent bottleneck in SoC development: for every new bridge instance, engineers hand-craft a full UVM testbench from scratch, consuming 40--100 engineering hours of bring-up effort. We present an agentic Python framework that generates complete, protocol-aware UVM testbenches for arbitrary IP and subsystem configurations from a structured Verification Manifest---a JSON specification encoding interface topology, protocol semantics, and coverage intent. The framework employs a dual closed-loop architecture: an inner compiler feedback loop that injects categorised UVM diagnostics rather than raw tool output, and an outer coverage loop driven by intent-gap analysis against the Manifest. Validated on the AMBA AHB-to-APB bridge---a representative production protocol bridge with a 7-state control FSM, two-phase APB handshake, and non-trivial burst serialisation logic---the framework achieves 96.8% functional coverage, a $4.7\times$ bring-up speedup, and a Protocol Verification Completeness score of 0.87, with only 2.1% of the generated 4,623-line testbench requiring human correction. The generated scoreboard independently identified a critical RTL functional fault that the manual baseline missed entirely. Generalisation across four protocol families confirms that speedup increases with protocol complexity, ranging from $3.1\times$ (APB Subordinate) to $6.1\times$ (CHI Subordinate).
Printed circuit board (PCB) routing is the stage of physical layout that finds paths connecting the placed pads of each net. Public datasets in this domain are few, and most contain only the final states of routed boards, not demonstrations of the electronic design automation (EDA) engine actions that produce them. To close this gap, we propose CADET (Constructing executAble Demonstrations for Engine-grounded Training), which jointly generates synthetic boards and demonstrations built from EDA engine actions. The resulting RouteDemos provides 100,000 synthetic boards routed with zero design rule violations (DRVs), together with their demonstrations. Their distributions over 17 statistics resemble those of 142 real reference boards. CADET trains a routing policy on these demonstrations with behavior cloning and fine-tunes it with reinforcement learning (RL). On real test boards, the policy raised the rate of complete DRV-free routing from 38.0% to 67.2% over an RL-only baseline. This rate was comparable to that of the open-source autorouter FreeRouting and exceeded those of the other baselines, including GPT-6 Astra. We release the dataset and its generation code as open source at https://github.com/LGAI-Research/CADET.
Photonic neural accelerators offer ultra-low latency linear algebra, yet their usable numerical precision, sustainable throughput, and application accuracy remain tightly bounded by surrounding electro-optical interfaces and physical non-idealities. Pre-fabrication design-space exploration (DSE) requires a cross-layer electronic design automation (EDA) framework capable of evaluating disparate optical topologies under identical neural workloads and physical assumptions. We present \textbf{PISim}, a unified Photonic Inference Simulator that couples pretrained deep neural network execution with physics-based models of six accelerator architecture families, six electro-optic modulator technologies, four optical source organizations, and direct or coherent readout schemes. PISim embeds stage-resolved impairment operators---covering control and analog-to-digital quantization, finite extinction ratio, shot noise, correlated relative-intensity noise (RIN), thermal drift, propagation loss, and channel crosstalk---alongside an analytical cycle-level interface timing model to expose electronic-optical bottlenecks. Using simulated MNIST multilayer perceptron (MLP) benchmarks, we demonstrate how PISim identifies optimal precision operating points under multi-impairment stress (peaking at 6 bits for Mach--Zehnder and microring modulators), quantifies a 52.5 percentage-point accuracy gap between VCSEL-array and parallel-fanout laser architectures arising, under the modeled conditions, from power partitioning and correlated source noise, and models interface-limited throughput scaling across emitter-array widths from 32 to 128 channels. This comprehensive framework establishes an extensible, reproducible foundation for accuracy-constrained Pareto optimization prior to photonic tape-out.
Users of artificial intelligence services, including those powered by large language models and other transformer-based models, need assurance that inference and training follow the declared computation without exposing private models or data. Zero-knowledge proofs can provide this assurance, but proof generation remains expensive. Hardware acceleration research often reports gains for individual cryptographic operations, leaving their effect on the full cost of proving neural-network workloads, particularly large transformer models, unclear. We introduce a four-layer framework linking verifiable AI workloads, proof systems, cryptographic kernels and hardware platforms, use it to review the literature, and add original measurements of six provers and seven kernels to identify which costs matter in practice. Our measurements show that the dominant cost changes with the proof system and input length. In low-rank fine-tuning, the cost of opening commitments to fixed model weights remains nearly constant across input lengths from 64 to 2048, while the work associated with intermediate values grows. A 13-billion-parameter fine-tuning configuration exceeds the capacity of a 40-gigabyte graphics processor but completes on an 80-gigabyte device, showing why capacity must be assessed alongside execution time. Our kernel profiling further identifies instruction issue and synchronization limits that a bandwidth-only model misses. We use these findings to assess graphics processors, reconfigurable logic, custom chips and existing artificial intelligence accelerators. The resulting research agenda connects proof-friendly numeric formats, reuse of fixed-weight commitments, heterogeneous execution and shared benchmarks to the measured bottlenecks. Effective acceleration requires selecting hardware and proof-system optimizations around the cost profile of the target workload.
Photonic neural network accelerators have emerged as a promising solution for reducing the power consumption of deep neural network (DNN) inference by executing matrix multiplication on analog photonic hardware. However, their inference accuracy is limited by low-bit quantization required for photonic processing and by analog photonic noise. This paper proposes an approach for improving inference accuracy by exploiting the high similarity among activation vectors in DNNs. Specifically, our approach constructs a representative vector for each group of similar activation vectors, computes its inner product digitally with high precision, and offloads the residual inner products to the photonic accelerator. Since the activation vectors within each group are highly similar, the residual vectors have much smaller norms, thereby reducing the effects of quantization errors and analog photonic noise. Furthermore, our approach preserves the photonic inner-product computation pipeline, allowing seamless integration into existing and future photonic accelerators without hardware modification. Experimental results on six representative DNN models demonstrate that the proposed approach can improve inference accuracy across diverse network architectures and application domains.
Reinforcement learning (RL) increasingly targets small language models, and executing these workloads efficiently across multiple accelerators requires effective coordination. On TPUs, Pathways-based RL uses a single controller to coordinate computation across hosts. This execution model is effective for large models, but its coordination computational costs are harder to amortize when small models repeatedly invoke short computations. To characterize these computational costs, we sweep dense models from 0.6B to 405B parameters on TPU v7x under a decode-heavy workload and measure how host and network computational overhead vary with model size. At 0.6B, host orchestration occupies 36% of step time and network communication 2%, whereas for large models, estimated network computational costs dominate the non-compute time. Motivated by these findings, we present Local Coordination JAX (LoCoJAX), which implements local rollout and reward processing on each TPU host with synchronous global training. In controlled comparisons against a tuned and optimized Pathways baseline on Qwen3-0.6B across three workloads and 8-64 TPU v7x chips, its colocated mode (LoCoJAX-CO) is 1.3-3.6x faster per step; on GSM8K, it completes a step faster on 8 chips than Pathways does on 64. Finally, a Ray-managed disaggregated mode (LoCoJAX-DA) overlaps rollout and training on separate hosts, and placement experiments establish when this overlap compensates for dividing devices and transferring weights: only when rollout dominates the step and receives most of the hosts. On NuminaMath-CoT at 256 chips, giving rollout 75% of hosts cuts step time by about 15% compared with LoCoJAX-CO.
Many hypotheses compete to explain the successes of knowledge distillation. To help address this, we propose and analyze a mathematical model of distillation, which suggests that distillation's performance comes not from obtaining better models but from easier to optimize landscapes. For generalized linear models trained with stochastic gradient descent, we prove that distillation fits performant student models asymptotically more quickly than non-distilled models. In rank-1 matrix approximation, we characterize conditions on the target matrix under which gradient descent with distillation converges strictly faster than training on the supervised objective. The theory helps delineate the ways distillation provides benefits (i.e., in optimization speed, not in generalization), and experiments on real datasets corroborate the theoretical predictions.
Deep neural networks for image processing are increasingly deployed on resource-constrained edge devices such as drones, medical systems, and smartphones, where limited computational resources and memory bandwidth pose significant challenges for efficient inference. Pruning during training often produces highly sparse convolutional kernels, offering the potential to significantly reduce computation and memory requirements. However, exploiting this sparsity efficiently at runtime remains challenging due to irregular memory access patterns and workload imbalance across processing elements. In this work, we investigate the standard convolution with sparse weights and analyze its performance characteristics through roofline analysis, showing that a 3 x 3 sparse convolution achieves an arithmetic intensity of approximately 10 FLOP/byte compared to roughly 96 FLOP/byte for its dense counterpart. We propose techniques for an efficient sparse convolution at high sparsities. Preliminary results demonstrate that our approach achieves upto 1.65x speedup as compared to dense convolution on NVIDIA A100 and upto 1.52x on NVIDIA Jetson Nano (Tegra X1) , motivating further exploration of direct convolution as a practical path to efficient sparse inference on modern accelerators/edge devices.
Merging independently fine-tuned checkpoints into a single multi-task model requires no data, no labels, and no gradient steps. In practice the task vectors of different tasks interfere, and the standard remedies--averaging, magnitude pruning, sign election--operate entry by entry, ignoring that each delta induces a low-dimensional update on the hidden features it acts upon. We recast merging as a subspace interference problem. For every two-dimensional weight tensor we take the top-r right-singular subspace of each task's delta--the input directions along which that task moves the representation most--and remove from each task vector the component lying in the union of the other tasks' subspaces. The resulting Joint Null-Space Projection (JNSP) is closed-form and order-independent, needs no forward pass, and is defined entirely by the checkpoint set. On eight CLIP image-classification tasks JNSP merges ViT-B/16 experts to 89.45% average accuracy without using a single image, 11.88 points above the strongest parameter-wise training-free baseline (TIES-Merging, 77.57%), 14.22 above Task Arithmetic, and 0.98 above WUDI's data-free consensus merge (88.47%). The gain holds across three backbones on the independently trained TALL release (+10.8, +12.6, +9.1 over TIES for ViT-B/16, ViT-B/32, ViT-L/14) and replicates there at 89.20%; controls confirm the source--a single shared subspace instead of per-task ones loses 2.9 points, and a sequential variant collapses to 69.6%. On RoBERTa-large the same projection underperforms plain task arithmetic, which we report rather than hide. Code is available at https://github.com.
General-purpose language models are trained mostly on English: over open models we review, at most, non-English pretraining data makes up only a 12% share. Models built for European languages give them far more, but score well below general-purpose models of the same size. A common concern is that, in small models, languages compete for capacity. We present Lara Mini base, a model trained from scratch on 6T tokens in English and 44 other languages, which together cover all 24 official languages of the European Union. As training progresses, the 44 other languages together come to outweigh English. To give these languages enough capacity at low cost, Lara Mini base uses a mixture of experts architecture with 7.2B total and 0.6B active parameters. Among open base models with up to 2B active parameters, it achieves the highest average score across standard English, multilingual, math, and code benchmarks, ahead of models trained on 4 to 6 times as many tokens. Its margin is largest on multilingual benchmarks in the languages it supports and on math benchmarks. It supports contexts of up to 128k tokens. We find that our MoE architecture reduces capacity competition between languages. For a single user on one GPU, without quantization, it generates text at least 1.8 times faster than any model with a higher average score. We release the weights of Lara Mini base under the CC BY-NC-SA 4.0 license. In addition, we will release T-HellaSwag, a higher-quality machine translation of the commonsense reasoning benchmark HellaSwag into 110 language variants.
Provably secure linguistic steganography hides a message inside text sampled from a language model, so that without the key, a stegotext is computationally indistinguishable from an honest model sample. Nearly every such scheme assumes the receiver can recompute the sender's token probabilities exactly bit for bit. We show this assumption is fragile: across three model families, bit-exact schemes decode almost none of our messages once the receiver's floating-point precision, quantization, or GPU architecture differs from the sender's, or even when the receiver merely scores a text in parallel versus one token at a time, with the same model, precision, and hardware held fixed. We present STEALTH, a scheme built for exactly this setting. The sender draws each token with an exponential race driven by keyed pseudorandom noise, so a receiver's probability mismatch flips a token only when the race is close; a rateless spinal code, decoded by beam search, absorbs the resulting flips without the sender and receiver ever agreeing on a single distribution. We prove stegotexts remain indistinguishable from honest samples under this construction. Empirically, STEALTH decodes every message we sent for every receiver in the sender's model family, at every precision and quantization level, on smaller models, and on a different GPU architecture. STEALTH is even robust across different model families with separate vocabularies and tokenizers, decoding every message we tested. Our approach carries far more payload per token than prior robust schemes, and trained steganalyzers and a targeted statistical test both fail to distinguish its stegotexts from honest model output.
We study multipoint-to-multipoint multicasting in synchronous radio networks with unknown participant locations and unrestricted message aggregation. Let $n$ be the network size, $k$ the number of participants, and $d$ their hop diameter. A supplied distance-two coloring with $C$ colors yields completion within $(C-1)d+1$ slots and at most $nk$ transmissions. In unit disk graphs, coordinates, one-hop neighbor data, and a common degree bound $Δ_0$ suffice for a $15(Δ_0+1)$-color schedule. A known fixed separation bound permits optimal deterministic $Θ(d)$ completion from coordinates alone. Without geometric data, strong selectivity gives $O(d\min\{n,(Δ_0+1)^2\log n\})$ slots when $n$ is known. Independent random transmission gives $O((Δ_0+1)(d+\log(k/η)))$ slots with failure probability at most $η$; for fixed $Δ_0$, this simplifies to $O(d+\log(1/η))$. A star construction shows that the logarithmic term is necessary for this random rule when the degree grows. We extend color schedules to fixed deletions and subgraphs connected over bounded windows of cycles. Separate conditional reductions preserve additive costs under diameter guessing and control parameter-table and cluster preprocessing. Finally, a counterexample shows that pruning and reranking a breadth-first tree can destroy interference-free delivery. For one-hop designated delivery with receiver degree at most $b$, we prove that the optimal schedule length equals the chromatic number of an associated conflict graph and is at most $2b-1$, with matching examples.
The growing volume of process data generated by screwdriving systems, combined with advancements in computing power, data storage, and algorithms, is facilitating the application of artificial intelligence in this field. Accurately detecting and classifying screwdriving errors is crucial for ensuring product quality, avoiding costly rework in manufacturing, and supporting sustainability by facilitating the maintenance, reuse, and recycling of screw-connected products. In this context, machine learning approaches offer a promising alternative to traditional monitoring techniques, such as window-based methods or envelope curves. This study examines the use of deep learning techniques for anomaly detection and error classification in screwdriving processes. The data corpus consists of multivariate time series capturing process parameters, such as torque and rotation angle. For anomaly detection, the embedding space of a vector quantized variational autoencoder was analyzed. In terms of error classification, four different model architectures were evaluated. The results demonstrate the effectiveness of anomaly detection and confirm that deep learning models can distinguish between multiple error categories beyond traditional binary classification. This extends the capabilities of conventional monitoring approaches, achieving accuracies of up to 100%. Additionally, real-world data material was examined in order to further demonstrate model robustness. Overall the results indicate strong generalization capability and highlight the practical potential of the developed models for use in industrial screwdriving processes, eliminating the need for expert knowledge in error identification and significantly reducing manual effort and analysis time.
Classical funnel control keeps the tracking error within a prescribed, time-varying envelope by increasing the feedback gain as the error approaches the boundary. For mechanical systems, this guarantee typically relies on velocity measurements. We show that it can be retained for fully actuated Euler--Lagrange systems using only position measurements. The proposed controller replaces the velocity error with a difference quotient of the position error over a fixed delay, while preserving the classical unbounded gain. It requires no plant model, reference derivatives, observer, or filter states. The central challenge is to reconcile the averaging inherent in delayed feedback with a gain that can grow without bound near the funnel boundary. We resolve this by establishing an a priori bound on the gain along closed-loop trajectories, independent of the delay. This bound yields a computable delay threshold that guarantees tracking within the prescribed funnel and boundedness of all closed-loop signals. The threshold depends only on bounds on the inertia, forces, and reference. Simulations demonstrate tracking accuracy comparable to full-state funnel control and tolerance of delays substantially longer than certified. With quantized position measurements, the controller also requires substantially less torque than an existing derivative-free funnel controller.
While quantization compresses LLMs and accelerates inference, outlier activations impede efficient low-bit representation. In particular, massive outliers cause substantial performance degradation. Although rotation-based methods that rely on randomized Hadamard transforms for online inference aim to mitigate this, they can produce fragmented distributions with empty bins when a massive outlier dominates a token's activation. This fragmentation severely undermines uniform quantization performance. Moreover, massive outliers concentrate in the down-projection input, where the preceding nonlinearity requires the rotation to be applied at inference time. This constraint prevents training-based approaches such as SpinQuant from learning the online rotation at this position, as a learned dense matrix cannot use the Fast Walsh-Hadamard Transform and would require a full matrix multiplication during inference. To address this, we propose SignRot (Sign-adjusted Rotation), which identifies massive outlier channels and constructs a sign adjustment vector that unifies the signs of the corresponding Hadamard row, collapsing the bimodal distribution into a unimodal profile amenable to quantization. Since the aligned outlier shifts the activation mean away from zero, we pair SignRot with asymmetric activation quantization, whose zero-point absorbs this shift. As a lightweight, drop-in improvement compatible with any Hadamard-based quantization method, SignRot requires only a single calibration forward pass and adds only element-wise sign flips at inference, without an additional matrix multiplication. Applying SignRot to QuaRot achieves a WikiText2 perplexity of 5.88 and an average zero-shot reasoning accuracy of 72.73, surpassing SpinQuant on the LLaMA3-70B model. Code is available at https://github.com/HYU-ARK-Lab/SignRot.
This monograph addresses the optimal selection of equipment for chemical-engineering systems of multiproduct batch plants, a large-scale NP-hard discrete optimization problem in which, for each processing stage, a unit of standard size must be chosen so that the total capital cost of the equipment is minimized while the system remains operable and produces the required products within a given time horizon. We present the mathematical model of system operation and analyze its computational properties. We then develop a sequential branch-and-bound algorithm and its parallel versions for shared-memory systems (OpenMP), distributed-memory clusters (MPI, master-worker), and graphics processing units (CUDA), including recursive and iterative tree traversal, shared-memory optimization, and branch-divergence reduction. To overcome the exponential growth of the run time, we propose a hybrid metaheuristic in which Ant Colony Optimization finds a feasible initial solution and Simulated Annealing improves it, and we implement it on GPUs with thousands of independent threads. Finally, we propose random, constant, and multi-constant approaches to tuning the parameters of the ant colony algorithm based on statistical analysis of many parallel runs, and we study the program on GPUs of three generations and on nodes with several GPUs. Computational experiments on multi-core processors, clusters, and GPUs show that the parallel hybrid algorithm finds solutions within 0.01% of the branch-and-bound optimum in tens of seconds instead of tens of hours.
Modern ML frameworks execute models using the kernel-per-operator paradigm, which runs each operator as a separate GPU kernel. This paradigm is modular, since developers can compose various kernel implementations and profile each kernel in isolation. However, this paradigm enforces dependencies at the granularity of entire kernels, which places a barrier between dependent kernels. These barriers cause streaming multiprocessors to idle, serialize computation and communication, and prevent a kernel from prefetching its weights while the preceding kernel runs. Megakernels remove kernel barriers by fusing the operators of an entire layer or model into a single kernel, but they give up the modularity of the kernel-per-operator paradigm. In this work, we ask whether a kernel-per-operator ML framework can perform on par with megakernels. Our key insight is that removing kernel barriers does not require fusing kernels. We develop MegaOps, a framework that launches dependent kernels on separate CUDA streams and enforces each dependency on individual tiles of the intermediate tensor that the producer kernel writes and the consumer kernel reads. Since kernels synchronize on data tiles instead of on each other's thread blocks, any two kernels that adopt the same data tiling can be composed without modifying either kernel. We evaluate MegaOps on dense Llama3 and Qwen3 models on single H100, A100, and L40 GPUs and on multi-GPU servers with 4$\times$A100 and 8$\times$H100. Our MegaOps framework matches the throughput of hand-tuned ThunderKittens megakernels, and achieves 1.09$\times$-1.10$\times$ the decode throughput and 0.98$\times$-1.10$\times$ the prefill throughput of the best-performing baseline, while reusing existing Triton, cuBLAS, and cuDNN kernels.
Expert-parallel Mixture-of-Experts decoding is latency bound: each layer serializes a token exchange, a grouped GEMM over routed experts, and a combine, while only a few tokens per rank are in flight. Prior GC-MoE fused these phases into one persistent, Green-Context-partitioned CUDA Graph on Hopper (SM90) and datacenter Blackwell (SM100). This paper rebuilds the operator for SM120 (compute capability 12.0, e.g. RTX PRO 5000) over PCIe. SM120 exposes block-scaled tensor cores only through warp-level mma.sync.aligned.kind::mxf8f6f4: no tensor memory, no two-CTA MMA, no cluster multicast. We rebuild the kernel around a swap-AB warp-level block-scaled MMA with FP4 weights in the register A operand and FP8 activations in B, plus SM120-specific mechanisms: scale-byte selection through MMA immediates, a compact-weights path, an adaptive token-tile geometry, an expert-granular streaming dispatch with a batched bulk-pull engine and a 64-bit deduplicating pull cache, full-width BF16 returns drained by resident relay copiers over a position-major arena, and shared experts folded into the compute role. On 8xRTX PRO 5000 over PCIe the operator is 1.522x and 1.387x faster than the FlashInfer CuTe DSL W4A8 MegaMoE and the Cake MegaMoE across 18 decode points (18/18 wins each), and 1.388x faster than Cake with shared experts fused; all 94 numerical cases and 752 per-rank checks pass, at cross-operator cosine similarity of at least 0.9999997.
Breathing monitoring is important for well-being, stress regulation, and guided breathing exercises, yet existing approaches often rely on contact-based sensors or controlled environments. This paper presents a non-invasive framework for respiratory phase detection using acoustic signals recorded on commodity smartphones. The proposed hybrid CNN-Transformer architecture combines local spectral feature extraction with long-range temporal modeling to capture the cyclic structure of breathing. The model is evaluated using a Leave-One-Subject-Out (LOSO) protocol on the Foundation Dataset and on an independent unseen test set comprising over 1300 breathing cycles and 2 hours of continuous recordings. The proposed 3-class formulation achieves 72.85% frame-level accuracy and 96.6% cycle-level exhalation detection efficiency in the LOSO evaluation, while achieving 76.94% frame-level accuracy, a Cohen's Kappa score of 0.65, and 99.2% cycle-level exhalation detection efficiency on the unseen test set. Results suggest that accurate respiratory monitoring may be achievable under data-constrained conditions, indicating that appropriate temporal modeling can learn transferable respiratory representations and generalize to previously unseen participants. An end-to-end ONNX deployment pipeline enables privacy-preserving on-device inference with latency below 20 ms, supporting real-time monitoring on mobile and wearable devices.
On-device inference requires kernels that make effective use of local accelerators. Efficient training and evaluation of models that generate Apple Metal kernels benefit from native macOS tools and concurrent agent sessions. These sessions alternate between model requests, development commands, and GPU execution, leaving GPU capacity idle for much of a rollout. We present Metal Sandbox, infrastructure that combines restricted per-instance file access, native command execution, and command-level scheduling across interconnected Macs. Each agent retains its workspace and private development environment. Parallel commands can run concurrently, while serial commands share one execution slot per node; eligible waiting commands can migrate to available Macs and return results to the original agent. Integration with Harbor supports task execution, automatic verification, and rollout collection through a common workflow. Controlled experiments on four Mac mini M4 nodes achieve $1.68\times$--$2.05\times$ gains in command throughput under concentrated demand. Sixteen static agent traces give an estimated $2.14\times$ sustained-throughput gain for four sessions per node versus one, assuming continuously replenished sessions and unchanged mean durations. Replaying these measured traces under dynamic scheduling reduces mean tool waiting by 18.8% and batch elapsed time by 22.3%, assuming unchanged phase durations. Metal Sandbox and its Harbor integration are open-source on GitHub.
Integrated electric vehicle chargers reduce cost by sharing power conversion hardware between the onboard charger (OBC) and the wireless charger (WC). A previously proposed dual active bridge (DAB)-based OBC/WC shares the rectification stage between the OBC and WC; however, the modulation required to realize its wireless operation has not yet been reported. This paper details a modulation strategy that enables the operation of the integrated OBC/WC topology. Differential-mode operation is used for the OBC, and common-mode operation is used for the WC. In OBC mode, the shared full-bridge produces differential-mode voltage for wired power transfer while ideally generating zero common-mode voltage on the wireless charging path. In WC mode, the same full-bridge produces common-mode voltage for wireless power transfer while ideally generating zero differential-mode voltage on the onboard charging path. The method preserves conventional DAB phase-shift power control and enables bidirectional power transfer in wired and wireless charging modes without additional contactors. Design considerations for the OBC transformer and WC coils are derived based on the resulting square-wave voltage amplitudes. Offline PLECS simulations demonstrate power tracking and decoupled charger operation at $\pm$7.4 kW. Hardware-in-the-loop simulations using a PLECS RT Box and an external Texas Instruments microcontroller further validate the proposed modulation method.
Recent music foundation models have achieved remarkable progress in music representation learning, yet most existing approaches rely on Transformer architectures whose quadratic attention complexity introduces substantial computational and memory overhead when modeling long-duration music recordings. In this work, we propose MusicRWKV, a multimodal music representation learning framework built upon the RWKV architecture. MusicRWKV jointly models waveform, spectrogram, and metadata information within a unified self-supervised learning framework. To improve efficiency for long musical sequences, we introduce an invertible compression module that reduces sequence length while preserving reconstruction capability. A FiLM-based conditioning strategy is employed to facilitate modality-aware representation learning without introducing expensive cross-attention operations. We further investigate the integration of Soft ROSA as a differentiable retrieval mechanism for neural-symbolic interaction. Experimental evaluation across a diverse collection of Music Information Retrieval (MIR) benchmarks demonstrates that the proposed RWKV-based architecture can be effectively adapted to heterogeneous music understanding tasks. A controlled pretraining ablation shows that Soft ROSA improves validation performance, while layer-wise telemetry analysis reveals a progressive transition from distributed retrieval to selective retrieval in deeper network layers. These results suggest that RWKV provides a promising architectural alternative to Transformer-based models for multimodal music understanding.
Reliable protection under varying power system operating conditions requires rapid fault detection while preserving selectivity against external faults and non-fault disturbances. This requirement is particularly challenging in communication-free schemes, where relays must determine trip conditions solely from local electrical measurements. To address this challenge, this paper proposes a directionally supervised neuromorphic protection scheme that integrates disturbance-responsive leaky integrate-and-fire dynamics with a direction-dependent firing threshold to distinguish internal faults from reverse-direction events. The unified relay architecture combines local disturbance detection, directional discrimination, and spike-based trip generation without requiring inter-relay communication. Its performance is evaluated under grid-connected and islanded conditions using a ring microgrid and the IEEE 13-node distribution feeder. Across 240 evaluated fault scenarios, the framework achieves median operating times of 5.4--7.6 ms across the two networks, with 78% of cases operating within 10 ms and 95% within 20 ms. It also maintains protection security during load disturbances, while fault-resistance analysis confirms its response under high-resistance faults. Controller-hardware-in-the-loop (CHIL) experiments using an RTDS simulator and an external dSPACE controller demonstrate trip-command assertion in approximately 10 ms and 5 ms for the tested single-line-to-ground and three-phase-to-ground faults, respectively. The results establish real-time feasibility within the evaluated conditions and characterize operating-time variability across fault types and resistance levels for different network configurations.
Protection of converter-dominated distribution networks is challenging due to limited fault currents, operating-mode transitions, bidirectional power flow, and heterogeneous network configurations. This paper proposes a scale-invariant neuromorphic protection framework to enable transferable protection decisions across such networks. The framework integrates asset-based normalization, multi-domain disturbance representation, event-driven spike encoding, leaky integrate-and-fire (LIF) dynamics, adaptive threshold regulation, and directional supervision within a unified mathematical formulation. Local voltage, current, active power, and reactive power measurements are transformed into dimensionless disturbance representations, while latency encoding and membrane integration capture disturbance intensity and temporal evolution. The framework is evaluated across four heterogeneous network configurations: a two-DER inverter-interfaced microgrid, an islanded ring microgrid, the IEEE 13-node feeder, and the CIGRE European medium-voltage benchmark. Across 480 cases, it achieves a median operating time of 7.6 ms, with 94.2 % of cases operating within one power cycle. The results demonstrate selective protection across different electrical ratings and network topologies using a common parameterization without network-specific retuning, highlighting the potential of neuromorphic processing for physically interpretable, communication-free protection of converter-dominated distribution networks.
Motivated by the potential of pneumatic excitable devices in soft robotics, this paper develops an energy-based, domain-agnostic equivalent of the Hodgkin-Huxley modelling framework using bond graph theory. The proposed model is used to guide the design of a pneumatic excitable neuron. The potential of this approach is illustrated through the design of a pneumatic half-center oscillator, an essential building block for rhythmic neuromorphic behaviours.
Repeatable test setups support the development and evaluation of integrated sensing and communications (ISAC) hardware and algorithms. Commercial radar target simulators are expensive, while sub-6 GHz ISAC experiments often need only tens of megahertz of bandwidth within their available spectrum allocation. This makes simpler, low-cost test hardware attractive. We present a low-cost, open-source radar target simulator based on an ADALM-Pluto without hardware modification, supporting four independently configurable targets and analog bandwidths up to 56 MHz. The FPGA generates echoes with independent delays, Doppler shifts and gains, using fractional-delay filters for sub-sample adjustment. For motion-consistent echoes, delay and carrier phase are updated together to reproduce both changing range and frequency-dependent Doppler shifts. Over-the-air measurements demonstrate simultaneous echoes, sub-sample delay adjustment, and smooth target motion between sample positions. The FPGA source, firmware and on-device web interface are publicly available.
Sampling methods are fundamental primitives for training and inference on probabilistic models. However, practical deployment of sampling algorithms such as Markov Chain Monte Carlo (MCMC) remains difficult on general-purpose processors. Existing hardware accelerators improve sampling efficiency, but they are typically specialized for a narrow class of models, graph topologies, or update schedules, making it difficult to support the diverse workloads used in real applications. This paper presents AMARA, an end-to-end compiler-driven accelerator framework for probabilistic sampling on arbitrary probabilistic graphical models. AMARA introduces MCMC-IR, a domain-specific intermediate representation that separates local factorized computation from global update dependencies. Based on this representation, the compiler performs architecture-aware partitioning, scheduling, and memory-access optimization to expose parallelism and reduce irregular graph communication. The optimized program is lowered to a compact ISA and executed on a programmable ASIC accelerator with specialized compute, sampling, and memory support. By co-designing the compiler and the architecture, AMARA bridges the gap between model generality and accelerator efficiency for practical MCMC-based probabilistic computing. We evaluate AMARA across representative workloads with different structures, including grid-structured and irregular energy-based models, and causal inference networks. The results show that AMARA supports a broad spectrum of discrete probabilistic models while achieving high sampling throughput and efficient hardware utilization. AMARA achieves an average of 6.1$\times$ (up to 16.3$\times$) higher sampling throughput than a JAX-GPU baseline on irregular workloads, and offers broader programmability than state-of-the-art accelerators.
Agent workflows issue dependent model calls that share long context prefixes, so serving them efficiently depends on reusing the key-value (KV) cache of those prefixes. On accelerators that compile model graphs ahead of time, however, the device KV budget is fixed, and prefixes evicted under load must be recomputed. Restoring them from host memory avoids this recomputation, but the transfer occupies the same engine thread as prefill and decode, so its benefit is not guaranteed. In this paper, we study when restoring beats recomputing on a commercial static-graph neural processing unit (NPU) stack (Rebellions ATOM+, vllm-rbln 0.10.2) that serves a Qwen3-8B planner on two cards and a Qwen3-0.6B executor on one. We first show that repeated prefill dominates planner work when default serving collapses: 89.5% of the executed planner prompt tokens had been materialized before. We then build a synchronous host-DRAM tier on the runtime's public KV copy calls, together with a scheduler-aware event model whose predictions are frozen before each test cohort. Experimental results show that the host tier raises workflow service-level objective (SLO) attainment from 22.8% to 100% and from 9.4% to 99.8% in the two tested overload conditions, and endpoint confirmation on two new seeds supports conservative load gains of 11.8% (chain) and 18.2% (fork), while mixed and bursty workloads remain seed-sensitive. The frozen model predicts median and 99th-percentile (P99) latency within its 15% tolerance in all 24 common-load conditions, including the 16-18% P99 increase that the tier causes for fork workflows, but it misorders three policy contrasts and overestimates the benefit of a cheaper store path (11% predicted against 3.1% measured). The trade-off therefore depends on runtime scheduling and transfer cost as well as cache capacity.
We present \parallelBatchBMatch, a parallel batch-vote algorithm for maximal $b$-matching in hypergraphs. The algorithm is deterministic for a given ordering of edges and produces the same output as the sequential greedy algorithm, i.e., the lexicographically first maximal $b$-matching. Under a uniformly random priority order it performs $O(m')$ expected work in the binary fork--join model, where $m'$ is the number of pins, the total number of vertex--edge incidences. The algorithm performs at most $m$ rounds for every priority order, where $m$ is the number of edges. Every round of the algorithm has $O(\log^2 m')$ depth with high probability, and thus the total depth of the algorithm is $O(m \log^2 m')$ with high probability. We implement the parallel algorithm in CUDA and evaluate it on real-world hypergraphs with capacities $b \in \{1, 3, 5\}$. The parallel algorithm computes the same matchings as the greedy algorithms and for the case of $b = 5$ it is up to $19.07$ times faster than the existing state-of-the art greedy algorithm.
LLM workloads sometimes ask many independent questions about the same long document, as in document QA, data labeling and LLM-based evaluation. Prefix caching removes the redundant prefill shared by these queries, but each decode step still reads the document's key-value (KV) cache once per question, so decoding is bound by memory bandwidth and scales with the number of questions. We present Parallel Prompting, which decodes P questions as P independent streams of a single request all at once, fixing this issue. In our method, every step reads the shared document once for all P streams, and a stream-id attention mask, evaluated inside the attention kernel, keeps each answer conditioned only on the document and its own question. We implement Parallel Prompting in SGLang by adding about 80 lines to three existing files and a small engine package; the mask needs no dense tensor, and every decode step has a fixed shape that is captured as a CUDA graph. With CodeLlama-7B, 8k-16k-token documents and 64-256 questions per document, Parallel Prompting is 3.7x faster than stock SGLang with its FlashInfer backend on an A100 and 5.9-6.0x faster on an RTX A6000, on two SGLang releases, including the most recent release. Timed on identical inputs, it is faster than stock SGLang in all 144 runs (2 GPUs x 2 releases x 6 workloads x 6 input sets). Code: https://github.com/jlab-nlp/parallel-prompting
Large language model (LLM) decode is dominated by short, predominantly memory-bound kernels that make bursty use of on-chip execution resources and memory bandwidth. This paper introduces SubGPU Pipelining, a fine-grained, intra-GPU spatial execution model that cuts an operator graph into ordered contiguous segments, binds those segments to spatial stages backed by disjoint streaming-multiprocessor (SM) pools, and streams microbatches through the resulting schedule. Selecting the graph cuts, segment-to-stage binding, per-stage SM allocation, and microbatch size is a coupled optimization problem: each choice changes the overlap pattern, shared-resource contention, cache state, and execution times of the others. We evaluate SubGPU Pipelining through both simulation and silicon experiment using the GPT-OSS 120B workload with 32-GPU B200 deployment settings. The analytical shared-resource simulator searches across SubGPU pipeline structures and batch sizes to map the per-user/per-GPU throughput Pareto frontier. At matched per-GPU throughput, the modeled SubGPU frontier provides up to 2.2x higher per-user throughput than temporal execution, and at matched per-user throughput it provides up to 1.9x higher per-GPU throughput. To test physical feasibility and further validate the simulation results, silicon experiments are performed, using one B200 GPU to emulate a 32-GPU B200 deployment. The experiment evaluates one selected two-stage pipeline configuration at global batch 2,048 using CUDA Green Contexts and TensorRT-LLM kernels. Results show that SubGPU Pipelining improves mean DRAM read-plus-write throughput from 71.8% to 83.6% of peak, and reduces the 36-layer decoder iteration time by 10.33%.
In recent years, large language models (LLMs) have become increasingly popular in software engineering and have performed well on different programming tasks. However, it is still difficult to know whether these models truly understand program behavior or whether they mainly rely on patterns learned from their training data. This distinction is particularly important for intermediate representations (IRs), where operations, dataflow, and execution semantics are explicitly represented. In this paper, we introduce MLIRBench, a benchmark for evaluating how LLMs reason about MLIR programs. MLIRBench uses a multi-agent pipeline to generate program mutations, validate their behavior through execution, and construct reasoning tasks from torch and linalg program pairs. The dataset contains 1,768 validated pairs, including 888 linalg pairs and 880 torch pairs, labeled as EQUIV, DIVERGE, or ERROR. We evaluated nine models, including five open-weight and four proprietary models. Gemini 3.1 Pro achieved the highest hard-label accuracy of 92.65% across all 1,768 pairs, counting unsuccessful predictions as incorrect. Across its 1,768 scored predictions, it achieved an aggregate score of 74.18 and a judge-assigned result-behavior score of 5.04/10. These results show that strong semantic classification does not necessarily imply accurate prediction of detailed execution behavior, highlighting a gap in current LLM reasoning over compiler IRs. By exposing this gap through execution-validated program pairs, MLIRBench provides a reproducible foundation for studying semantic reasoning over compiler IRs and related tasks such as program repair and semantic classification.
We present an open-architecture bidirectional neural interface for closed-loop neuromodulation built around a dual-core STM32H747 microcontroller and commercial Intan recording and stimulation front-ends. The system supports up to 128 recording channels at 20 kSps and 64 stimulation channels while enabling low-latency on-device AI inference through a dual-core processing architecture and a two-stream memory hierarchy. To validate the machine learning pipeline, we demonstrate two neural decoding models, a compact multilayer perceptron executing entirely from on-chip memory and a larger temporal neural decoder utilizing external memory staging. Both models successfully decode fingertip velocity from neural spiking activity. As a proof of concept for closed-loop operation, we implement spike-triggered stimulation and achieve an end-to-end latency of $189 \pm 14\,μ$s, representing an improvement of three orders of magnitude over a host-PC-mediated closed-loop pathway. These results demonstrate that resource-constrained embedded platforms can support high-channel-count neural interfacing, flexible on-device AI, and sub-millisecond closed-loop neuromodulation.
Acoustic Anomaly Detection (AAD) serves as a non-contact approach for machine condition monitoring, yet much of the existing research is evaluated offline and rarely addresses embedded hardware constraints or microphone variations under operational drift. This paper presents a comparative evaluation of on-edge continual AAD across two embedded platforms the ESP32-S3 (with an external I2S digital microphone) and the Arduino Nicla Vision (with an onboard Pulse Density Modulation (PDM) microphone) using a shared detection model. The model utilizes a lightweight ShallowNet encoder and contrastive prototype learning to map spectrograms into a 32-dimensional embedding space. Both boards run an identical TensorFlow Lite for Microcontrollers (TFLM) encoder, performing on-device prototype learning updated via an exponential moving average. A cloud pipeline orchestrated through Firebase and MQTT enables remote data collection and retraining. Evaluated via a controlled speaker-replay protocol using standardized Detection and Classification of Acoustic Scenes and Events (DCASE) Challenge Task 2 machine sounds, the ESP32-S3 achieved 82.35% to 95.00% accuracy across five machine types, while the Nicla Vision achieved 79.41% to 95.00%. The two platforms demonstrated comparable performance, confirming that the lightweight prototypical architecture transfers reliably across distinct hardware and sensor tiers.
AI models of the power grid must capture how wind and solar plants support the voltage, adjusting their reactive power to the local voltage along a grid-code Q(V) characteristic. This research examines whether grid foundation models, fast neural network models for power-flow analysis, can represent this behaviour. We introduce the Q(V) bus, a new bus type for the open-source GridFM GraphKit, whose physics decoder evaluates the reactive injection from the plant characteristic at the predicted voltage in every correction step. Training data comprise 7882 hourly operating points of a modified IEEE 118-bus system, matched to Energinet DK1/DK2 measurements and solved with pandapower. Compared with a physics-informed baseline that represents the parks as negative load, the Q(V) bus lowers the voltage-magnitude RMSE 2.3-fold, reduces the power-balance mismatch at the parks from 25 to 0.66 MVA, and raises the F1 score for voltage-band violations from 0.40 to 0.76. An ablation over five configurations shows that the Q(V) bus type provides the accuracy and the physics loss provides the power balance. For edge monitoring, we distil the trained $2.0\times10^{7}$-parameter model into students of $6\times10^{5}$ and $1.3\times10^{6}$ parameters; a flat-start initialisation prevents their collapse into an incorrect low-voltage state in which the characteristic saturates. With output distillation, the 12-step student reaches 1.4 times the teacher's error and, for edge inference on an NVIDIA Jetson Orin Nano (15 W), solves an hourly scenario in 12.2 ms, within the 20 ms synchrophasor interval.
Deployment of retrieval-augmented generation (RAG) for Edge-AI is constrained by memory and bandwidth limitations. A standard 7B parameter model consumes 14 GB Video Random Access Memory (VRAM) in half-precision (FP16), precluding co-location of vector indices and Key-Value (KV) caches. This paper presents an engineering integration and empirical evaluation of a multimodal RAG framework optimized for consumer-grade Turing hardware (e.g., NVIDIA Tesla T4). Our system-level analysis identifies a critical trade-off: while generative pre-trained transformer quantization (GPTQ) offers superior throughput (51.6 tok/s), it exhibits significant context instability at 2048+ tokens on Turing architectures. Conversely, NormalFloat 4-bit (NF4) maintains robust long-context performance with viable latency (81 ms/tok). By coupling NF4 with a decoupled Florence-2 vision integration, we achieve visual reasoning with only 430 MB VRAM overhead. Furthermore, empirical observations of prompt-induced search space pruning show that structured prompting reduces latency by 47% without altering computational intensity (1.88 GFLOPs/token). The integrated architecture maintains an 8.37 GB footprint, establishing a viable RAG pipeline for 12-16 GB GPUs. Finally, structured few-shot prompting effectively compensates for quantization-induced reasoning loss, restoring mathematical accuracy from 35.0% to 92.0% across 200 trials.
Can the functional architecture of a Transformer attention head be given a biological construction? More specifically, could a suitable group of pyramidal columns in posterior temporal language regions (PTLRs) constitute the biological construction of a neuromorphic mechanical realization of an attention head, under a mathematical instantiation that need not coincide with its standard Transformer instantiation? This paper investigates this hypothesis by examining possible correspondences between the functional organization of attention heads and neurophysiological mechanisms operating within and across pyramidal columns involved in human language processing. Our analysis builds on a mechanical-structural description of Transformer attention heads as systems of simple neural networks linked through output-input and output-weight interconnections. We examine the functional organization of an attention head and the input-dependent transformations it produces, and ask how the corresponding processing relations could be constructed in biological neural tissue. We then identify candidate neural mechanisms of signal transmission and modulation within and across pyramidal columns that may jointly support such a construction. On this basis, we propose that a suitably organized group of pyramidal columns in PTLRs could constitute a biological construction of a neuromorphic mechanical realization of an attention head. Whether the required functional organization actually exists in PTLRs remains an open empirical question. The hypothesis does not require the biological system to reproduce the standard mathematical instantiation of Transformer attention: the same functional architecture may be mechanically realized and mathematically instantiated in a different way, consistent with its biological construction.
Probabilistic-bit (p-bit) invertible logic encodes a Boolean relation as an energy whose ground states are its legal assignments, computing in any direction by sampling. Traditional constructions target ground-state membership and a wide minimum energy gap, yet finite-time samplers traverse excited states. We cast Hamiltonian design as inverse statistical mechanics: fit an energy with interactions up to order four whose Gibbs law approaches uniformity on each clamp case's legal states, under a gap floor keeping every legal state below every illegal one. Constrained maximum likelihood with pruning and refitting fits terms from a fixed pool, and every delivered energy is audited exactly on all cases. Each energy is scored by time to a verified solution under restarts with its best of nine tuned samplers. In underdetermined modes, learned multipliers factor in 0.55, 0.44, 0.52 and 0.46 [0.38, 0.56] of the hand-written encodings' updates from 4x4 to 7x7, faster within every family below 7x7 at no resolved work penalty; learned adders solve the reverse mode in 0.71 of the updates at 8 bits and about a quarter at 24 and 32 bits, at 0.55 of the work at 32 bits. At 5x5 under equal pool, budget and floor, a gap-maximizing fit reaches 0.79 [0.73, 0.88] in updates and 1.26 in work against the hand-written encoding, while likelihood fitting takes 0.55 [0.46, 0.58] and 0.58 [0.48, 0.61] of the gap fit's updates and work. Rounding coefficients to the narrowest exact integer words tested, we run both energies at each pair's common word on one FPGA color-sweep SSA core, bit-exact with software: learned multipliers take 0.41 of the hand-written encoding's cycles at 4x4 and 0.057 [0.045, 0.068] at 7x7, at up to 2.4 times its lookup tables. The method requires only a circuit's legal assignments; in determined modes, hand-written encodings take less work, and from 12 bits on the adder fewer updates.
DNN surrogates are increasingly used to accelerate scientific workflows that must process high-rate sensor streams in near real time, exemplified by Ptychographic imaging at next-generation X-ray light sources. Such surrogates deliver 100-1000x higher throughput than iterative solvers, enabling faster steering of experiments. To preserve surrogate accuracy during an experiment, as the input distribution evolves, state-of-the-art deployments freeze the pretrained surrogate or run a separate CUDA context that fine-tunes concurrently with serving. However, multi-context concurrency causes significant contention on the GPU, inflating both the mean and the variance of inference latency, triggering deadline misses in workflows where sensor data has to be processed within a hard per-unit deadline. To address this challenge, we propose UniPipe, a single-context system that unifies inference serving and fine-tuning of a DNN surrogate on a single edge GPU. UniPipe is driven by three key insights: both serving and fine-tuning require a forward pass and can be fused into one; the per-sample latency of modern GPUs is non-linear in batch size; and the workload is repetitive across intervals, enabling offline scheduling. Specifically, UniPipe fuses training and inference data in a single forward pass with a selective backward pass on the training subset, and uses a polynomial-time scheduler that exploits the latency curve to compute the batch composition. We evaluate UniPipe on a real-world Ptychographic reconstruction workflow with sensor rates up to 5000 patterns/sec.; UniPipe lowers reconstruction MSE by up to 30% and inference MSE by up to 50% over a pretrained baseline, and reduces the deadline-miss rate by 13%-22% over a multi-context baseline. UniPipe's unified deployment reduces the latency spread from 106% to 57% of the mean, and increases total throughput compared to multi-context baseline.
Genetic programming (GP) evolves populations of heterogeneous programs, whereas GPUs achieve their highest throughput when groups of threads execute the same instructions. To better align GP evaluation with GPU execution, we present Galop, a hybrid CPU-GPU linear GP system that assigns one individual to each CUDA warp, a group of 32 threads that execute the same instructions across different fitness cases. We compare Galop against EvoGP, a recent high-performance GPU-based tree GP system, on data-rich symbolic regression across two unique GPUs, varying population size and fitness-case count. Galop completes evolutionary runs faster than EvoGP in 27 of the 28 main comparisons, with speedups of up to 13.6 times. Solution quality comparisons favour EvoGP overall, although Galop remains competitive on several problems. Galop also completes a full generation with 2.15 billion individuals on a single H100 GPU, approximately 4.0 times the largest population achieved by any other tested GP system. All five runs at this population complete 100 generations, with a median wall-clock time of 2.23 hours. These results make it practical to study GP populations in the billions on a single GPU. Galop's source code, experimental data, and reproducibility scripts are publicly available at https://github.com/xavierhillroy/Galop.
Training large language models to generate efficient CUDA kernels requires supervision spanning diverse computational requirements and iterative code optimization. Constructing this supervision is challenging because of heterogeneous specifications, and useful implementations emerge through iterative complicated testing and refinement. We introduce KernelMosaic, an open dataset pairing diverse CUDA tasks with verified solutions and trajectories. Human-guided review and diversity-aware selection curate tasks expressed as natural-language descriptions, reference programs, mathematical specifications, and their intersections. To build the solutions and trajectories, our agentic framework guides a strong LLM through implementation, correctness repair, and performance optimization with reusable CUDA guidance and execution feedback. Supervised fine-tuning of a small open-source model on KernelMosaic raises correctness from 4%, 0%, and 0% to 88%, 82%, and 16% on KernelBench L1, L2, and L3, respectively, while increasing the fraction of tasks with kernels faster than PyTorch eager at every level. Further training experiments show that supervision from the curated trajectories substantially outperforms direct responses and promotes sustained testing and revision. These results demonstrate that KernelMosaic enables a general-purpose model to achieve substantial gains in CUDA kernel generation.
Second-order optimizers improve deep neural network training using curvature information, but their practical efficiency is limited by GPU execution overhead. AdaFisher reduces Fisher Information Matrix (FIM) approximation costs through diagonal Kronecker structures and exponential moving averages, yet its PyTorch implementation requires sequential CUDA kernel launches, intermediate tensor allocations, and repeated memory accesses. We introduce CAFIN, a GPU-native CUDA kernel fusion framework that improves AdaFisher execution efficiency without changing its optimization formulation. The proposed approach progressively integrates FIM preconditioner construction, exponential moving average updates, gradient preconditioning, momentum accumulation, weight decay, and parameter updates into fused CUDA kernels. Shared-memory reuse, register-resident computation, and in-place operations further reduce memory traffic and intermediate allocations. Experiments using ResNet-18 on CIFAR-10, CIFAR-100, and Tiny ImageNet across NVIDIA GTX 1660 Ti and RTX 3060 Laptop GPUs demonstrate up to 78.4% lower isolated FIM construction latency and 2.1-26.9% lower optimizer-step latency. The implementation also eliminates approximately 18.91 MB of intermediate GPU memory allocations per optimizer step. Numerical validation confirms consistency within the predefined FP32 tolerance, while convergence accuracy remains comparable across evaluated datasets and random seeds. These findings demonstrate that CUDA kernel fusion can substantially accelerate second-order optimization while preserving AdaFisher's underlying mathematical formulation and learning behavior.
Wireless edge inference enables Physical-AI devices to execute edge-inferred actions using transmitted sensory inputs. Since the risk of incorrect actions vary across observations and tasks, the associated uplink transmission of sensory inputs requires different levels of wireless reliability. For instance, applying a fixed signal-to-interference-plus-noise ratio (SINR) requirement can overprovision low-risk tasks while underprovisioning high-risk ones. Therefore, we propose risk-aware control for edge inference (RACE), which maps predicted action-failure probability and task criticality to adaptive SINR targets and combines them with deadline-derived SINR requirements. A slow-timescale large language model (LLM) proposes task requirements for deterministic validation, while a fast hybrid actor selects inference configurations and transmit power levels, complemented by analytical receive beamforming and finite power restoration. Simulations show $11\%$ improvement in the average accuracy-latency-energy objective over fixed-high-SINR control, with near-zero constraint outage.
Bayesian inference weighs competing explanations of noisy data by sampling from a posterior distribution, a task that deterministic hardware performs only by simulating randomness. Probabilistic bits (p-bits) supply this randomness physically, but the few hardware networks of p-bits built so far use inorganic devices, address optimization or learning, and are judged without reference to the spread expected from finite sampling. Here we report a probabilistic computer built from solution-processed conjugated-polymer memristors. Four polymer p-bits, tracked online for drift and coupled in a closed Gibbs-sampling loop, sample competing, degenerate and cooperative posteriors over sixteen states with statistics that fall within the range of ideal p-bits, both in their distribution and in their autocorrelation, without any parameter fitted to the target. This agreement extends to new simulated sensor readings whose outcome was predicted before each run, including readings that point away from the true answer. Reprogrammed in software, the same circuit and controller reproduces the invertible AND gate of earlier p-bit networks, with its temperature fixed in advance rather than fitted. Polymer p-bits thus open a route to printed, large-area probabilistic computers that infer alongside their sensors.
Graph learning on neuromorphic substrates requires an explicit spiking realization of neighborhood aggregation, supporting control and classification. We propose DeNA-SGN, which combines fixed OR and tie-inclusive Majority aggregation of binary node features with a learned signed-weight threshold classifier. A degree-dependent inhibitory construction implements Majority through shared control neurons, enabling a complete circuit compilation into SuperNeuroMAT. Across seven datasets and 41 mapped classifiers, native simulation matches the exported ANN on every evaluated class decision and output-spike pattern. On the new confirmatory Amazon Photo dataset, DeNA-SGN achieves 83.93+-0.93% current-readout accuracy and 71.61+-2.22% strict output-spike accuracy; Wiki-CS reaches 75.99+-0.54% over 20 public masks. In the matched Cora study, we achieve 74.90+-2.67% on current-readout and 68.57+-2.94% strict spike accuracy. Complete circuit and activity accounting supports analytical energy comparisons and identifies the cost of clocking inactive neurons in large deployments. Current argmax is decoded on the host, and energy is modeled rather than measured on hardware. These findings establish a reproducible route to useful graph classification with decision-matched spiking execution and explicit implementation costs.
Near-data processing accelerates data-intensive applications by moving computation closer to data, reducing data movement overhead and improving performance and efficiency. This paradigm is particularly promising in genome analysis, where read mapping, a critical step for aligning DNA sequences to reference genomes, poses substantial challenges due to the sheer volume and complexity of the data. Traditional software solutions and high-performance computing (HPC) clusters struggle with the rapid growth of genomic datasets, while specialized hardware accelerators remain limited by data movement. In this paper, we present Mendel, a heterogeneous near-storage architecture for end-to-end long-read mapping that combines storage-side, memory-side, and FPGA-based acceleration within an SSD-centered design. Mendel features a superscalar chunk-based pipeline and a chunk-aware filtering heuristic, specifically designed to handle long DNA sequences. Under our modeling-, simulation-, and synthesis-based evaluation, Mendel outperforms the long-read hardware mapper Darwin by up to 2.97<inline-formula><tex-math notation="LaTeX">${\boldsymbol{\times}}$</tex-math><alternatives><mml:math><mml:mrow><mml:mo mathvariant="bold">×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="mercogliano-ieq1-3733176.gif"/></alternatives></inline-formula> in performance and 17.97<inline-formula><tex-math notation="LaTeX">${\boldsymbol{\times}}$</tex-math><alternatives><mml:math><mml:mrow><mml:mo mathvariant="bold">×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="mercogliano-ieq2-3733176.gif"/></alternatives></inline-formula> in energy efficiency. Compared with Darwin enhanced with GenStore-style in-storage filtering, Mendel also provides higher projected performance and lower projected energy consumption in our evaluated setting.
With the growing interest in using AI (Artificial Intelligence) for RTL (Register-Transfer Level) hardware development, robust and comprehensive verification has become more important than ever. As Large Language Models (LLMs) increasingly assist in creating and modifying RTL designs, ensuring that these changes preserve existing functionality is paramount for building trust in AI-driven workflows. We present SVApshot, a fully automated framework that leverages LLMs to generate and systematically attempt to correct SystemVerilog Assertions (SVA) for RTL modules, completely removing the human from the generation loop. SVApshot introduces a novel snapshot methodology that captures the current functionality of a design as a comprehensive set of formal assertions, creating a regression suite for validating future changes--whether made by humans or AI. The framework features automated duplicate detection, assertion set expansion, and iterative repair of failing assertions using LLM-guided debugging with formal verification feedback. Our experimental evaluation across diverse RTL modules demonstrates high coverage for complex modules, with successful detection of both manually-injected and AI-introduced bugs. Additionally, this framework opens the door for automatic generation of formal testbenches to benchmark LLMs for hardware design, moving beyond binary pass/fail metrics towards nuanced assertion-based scoring.
Face-to-face (F2F) hybrid bonds shorten interdie connections, but data, clock, and power-grid (PG) connections must be realized together. CU-3D is a RTL-to-GDS physical design flow based on commercial EDA tools. It derives role-compatible hybrid-bond (HB) sites from a 17-metal ASAP7-based technology model, complete HB footprints, and constructed PG geometry. To coordinate partitioning with placement flexibility, Scaled refines an inherited 2D placement, whereas Fresh places the receiving die first and uses its pin locations and clock loads to guide placement of the other die. Data and clock HB assignments are jointly optimized while retaining alternative data sites and reserving capacity for subsequent PG assignment. The selected HB sites and stacks are preserved during 3D clock-tree synthesis, supply-aware cell insertion, and routing. Across eleven designs, including a 133-macro processor, the flow achieves zero backend DRC violations and passes signal-connectivity checks. Nine paired designs admit lower-footprint, lower-power implementations; four improve setup WNS. On JPEG, early joint planning gains 87 ps and mobility-aware partitioning gains 346 ps. Path analysis explains why cell-delay savings can outweigh net-delay increases.
High-level synthesis (HLS) can translate C/C++ kernels into register-transfer level (RTL) designs, yet generating high-quality hardware still relies heavily on manual pragma tuning and repeated synthesis evaluation. We propose AutoC2V, a C++-to-Verilog generation framework driven by real synthesis feedback that enables large language models (LLMs) to directly generate RTL with improved quality of results (QoR) from synthesizable C++ kernels. We construct training data from real synthesis results, including synthesizable C++ source code, the corresponding Verilog RTL, QoR metrics, and optimization objectives. Using this dataset, we align the model toward synthesis-aware RTL generation through Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO). To reduce the high cost of repeatedly invoking real synthesis during GRPO, we design the QoR Prediction Reward Model (QoR RM), which predicts latency and resource metrics by encoding RTL into long-context semantic representations, thereby providing low-cost and continuous QoR reward signals for policy optimization. Experimental results demonstrate that AutoC2V not only substantially improves RTL functional pass rates, but also achieves superior latency and resource optimization compared to the unfine-tuned base model, representative HLS optimization approaches (HLSPilot, LLM-DSE, and AutoDSE), and advanced general-purpose commercial models (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, DeepSeek v4 Pro, and GLM-5.2). These results indicate that combining real synthesis data with efficient QoR reward modeling can further improve the ability of LLMs to generate high-quality RTL.
The following paper discusses a test bench to measure the consumed power of a hardware accelerator module implemented in Verilog. The AMD’s Vitis High-Level Synthesis (HLS) integrated development environment (IDE) has been used to synthesize the module. The final design is validated on a demo board and measured energy is compared to AMD’s Vivado estimates.
Reliability in safety-critical systems is essential for operation in environments prone to ionizing radiation. This work presents the design and reliability analysis of a hardened, 5-stage pipelined RISC-V processor integrated with a configurable and fault-tolerant instruction memory cache. To balance performance and dependability, hardening strategies, including Triple Modular Redundancy (TMR) and Error-Correcting Codes (ECC), are selectively applied to both the pipeline datapath and the cache structures, complemented by built-in observability mechanisms to track real-time error propagation. System-on-Chip (SoC) simulation results demonstrate a peak performance speedup of $3.38\times$ with the integrated cache configurations. Physical synthesis on a SmartFusion2 M2S025 FPGA highlights the hardware trade-offs, showing a 41% LUTs and 43% DFFs overhead for the hardened core with cache while maintaining a compact footprint. Finally, comprehensive fault-injection campaigns under simulated neutron flux reveal that while the largest baseline cache increases failure rates by $3.4\times$, the proposed hardening architecture achieves an average reduction in error propagation of up to 97%, restoring the system's failure cross-section to levels comparable to those of the hardened core without cache.
Training neural networks directly on edge hardware is useful when a system must adapt to local data while avoiding continuous cloud dependence. This paper describes the registertransfer-level (RTL) design and simulation verification of a $16 \times 16$ fixed-point t wo-layer neural network trained with the Adam optimizer. The design is written in Verilog HDL and verified using Xilinx Vivado XSim. All major datapath operations use signed 32-bit Q16.16 fixed-point arithmetic. A $16 \times 16$ diagonal-dataflow systolic array is modeled at RTL as the matrix-multiplication engine for both forward propagation and backpropagation. The Adam update path is described as an RTL datapath using an integer four-iteration Newton-Raphson square-root unit and a 256-entry bias-correction look-up table. Verification is performed against a Python float64 g olden m odel o n f our l earning tasks: identity mapping, diagonal binary classification, bipolar identity mapping, and XOR classification. T he i nitial loss a grees within 0.05% across the testbenches. The RTL obtains 100% accuracy on the diagonal classifier and X OR t ests, a nd the X OR meansquared error (MSE) is reduced by 81.03% over 250 training iterations. These results validate a complete fixed-point RTL training engine built around the proposed diagonal systolicarray core. The scope of this work is simulation verification; FPGA synthesis, implementation, timing, power, and resourceutilization results for the training engine itself are discussed as expected outcomes, using a companion hardware-validated implementation of the same array as an evidence-based proxy, and are left as the immediate next step of this work.
This work uses the Python language as a modeling tool to implement a GFSK demodulator for Bluetooth Low Energy. Firstly, a classic modulator and a non-coherent demodulator based on FM discriminator architecture are modeled in Python. Secondly, a comparison of demodulators using a bit-error-rate metric is performed to define the target sampling rate, achieving 19 dB SNR for a 0.1% BER using 16 samples per symbol. Finally, both blocks are implemented in Verilog using the Python model as a reference. The physical implementation of the 16-MHz Verilog modulator and demodulator, synthesized for a 28 nm Digital Process Design Kit resulted in areas of $1200 \mu m^{2}$ and $2500 \mu m^{2}$, respectively.
Graph Convolutional Networks (GCNs) are widely adopted for tasks involving relational or graph-structured data and can be formulated as two-stage sparse–dense matrix multiplication (SpMM) during inference. However, existing accelerators often struggle with the irregular workloads induced by power-law node degree distributions. In this work, we propose <italic>FlexVector</italic>, a domain-specific vector accelerator that efficiently executes SpMM for GCN inference under irregular workloads. FlexVector adopts a row-wise, product-based dataflow that regularizes SpMM execution and exposes vector parallelism through full-row access to vector registers, eliminating the need for multi-banked register file designs. Building on this dataflow, it introduces software-managed, flexible vector register files (VRFs) that adapt to irregular data access patterns, without sacrificing memory access efficiency. To further exploit these architectural capabilities, we develop a graph preprocessing strategy that restructures irregular graph workloads to better match the row-wise dataflow and VRF capacity. This hardware–software co-design reduces memory traffic, leading to significant performance and energy efficiency gains on real-world GCN workloads. Experimental results on five real-world GCN datasets show that the VRF-centric FlexVector achieves a <inline-formula><tex-math notation="LaTeX">$\mathbf{3.78\times}$</tex-math><alternatives><mml:math><mml:mrow><mml:mn mathvariant="bold">3.78</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="yang-ieq1-3742755.gif"/></alternatives></inline-formula> speedup and 40.5% lower energy at comparable area cost relative to a state-of-the-art cache-centric baseline with buffers of the same size.
Purpose. The aim of the study is to systematize and comparatively analyze methods of decomposition and modeling of microservice architecture for forecasting systems, as well as to substantiate promising directions for the development of this research area with regard to the specific characteristics of analytical workloads. Methodology. The methodological basis is a systemic and comparative analysis of decomposition methods: classification of approaches into static, dynamic, hybrid, and artifact-driven; comparison of graph theory methods, Domain-Driven Design (DDD), and artificial intelligence (NLP, graph neural networks) according to criteria of automation level, domain semantics coverage, and suitability for ML components; formalization of workload heterogeneity. Scientific novelty. Based on a generalization of MLOps practices, the key types of heterogeneity in the analytical workloads are formalized as criteria for decomposition effectiveness: compute asymmetry between ML model training and inference, a dual-loop data lifecycle structure (stateful/stateless), and the isolation of technological dependencies of components. A hybrid three-stage approach is proposed, combining DDD to fix logical domain boundaries, dataflow-oriented optimization to isolate data loops, and automated graph-based and AI methods for the verification of microservice boundaries. Conclusions. It is established that decomposition methods form four classes, among which hybrid approaches demonstrate the best quality yet remain the least common. It is proven that methods with the highest degree of automation (graph-based, AI) are insufficiently effective in terms of compute asymmetry and stateful/stateless separation. Solving the problem of the absence of decomposition methods that account for workload specifics lies in developing hybrid approaches with dynamic verification based on spatiotemporal graph neural networks.
Deep Neural Networks (DNNs) are increasingly deployed in safety-critical systems, where real-time inference is commonly accelerated through dedicated architectures such as Systolic Arrays (SAs). Nevertheless, in such accelerators, hardware faults may alter inference outcomes, potentially leading to sample misclassifications. Consequently, regulations require self-testing mechanisms capable of detecting faults and identifying their location to enable appropriate mitigation strategies. However, exhaustive hardware testing solutions, such as Built-In Self-Test, typically require dedicated operating modes, resulting in service disruption and increased energy consumption. To address these limitations, in this work we propose a test-pattern reduction strategy tailored to SAs. In particular, the proposed approach retains only the test vectors covering the most impactful fault locations by first identifying them through fault injection and subsequently selecting the patterns required to ensure their coverage. We validated this approach on LeNet-5 and ResNet-8 models mapped onto output-stationary SAs, showing a test-pattern reductions of 63.2% and 64.4%, respectively.
Systolic Arrays (SAs) have become a widely adopted architectural paradigm in modern Artificial Intelligence (AI) accelerators. However, ensuring their reliability through manufacturing test introduces significant challenges. Conventional Design-for-Testability (DfT) methods rely on invasive scan-chain insertion, leading to increased silicon area and excessive test time overhead. Consequently, SA architectures require efficient and low-overhead test strategies. This work proposes a minimal-overhead DfT strategy that leverages a key property of systolic architectures, namely the intrinsic cascading of functional registers across the Processing Elements (PEs), naturally forming shift-register structures. The proposed solution is versatile, supporting both Weight Stationary (WS) and Output Stationary (OS) dataflows. By exploiting these native register-chains, test patterns are applied through the functional datapath, effectively reducing the array-level test complexity to PE-level testing. Furthermore, this approach enables parallel testing of PEs and supports finegrained fault diagnosis for the exact localization of faulty PEs.
Large Language Models (LLMs) exhibit extreme compute and memory demands, necessitating large-scale multi-chiplet systems to achieve high-performance and scalability. In such systems, frequent memory accesses for model parameters and growing KV cache induce substantial data movement between compute and storage chiplets, making inter-chiplet communication a critical performance and energy bottleneck. Silicon photonic (SiPh) Networks-on-Interposers (NoIs) enable high-bandwidth and low-latency communication. However, SiPh integration introduces significant packaging challenges. Thermal management becomes critical as SiPh-based devices are highly sensitive to temperature variations. In this paper, we highlight the design principles of a scalable, dataflow- and thermal-aware NoI architecture tailormade for LLM inference.
As Human-Machine Interaction (HMI) requirements evolve in modern System-on-Chip (SoC) architectures, high-efficiency synchronous serial communication has become important for data exchange. However, traditional software-defined Serial Peripheral Interface (SPI) drivers in high-performance embedded systems often impose significant Central Processing Unit (CPU) overhead and timing jitter, as the processor must manage byte-level transfers. To address these limitations, this paper presents an Field Programmable Gate Array (FPGA)-based SPI hardware accelerator that offloads data movement and protocol management from the CPU. Optimized via heterogeneous storage architecture, the Intellectual Property (IP) core utilizes a 2048-depth 9-bit synchronous First In First Out (FIFO) to achieve atomic encapsulation and streaming decoupling. At the logic layer, a three-state Mealy Finite State Machine (FSM) is developed to eliminate startup latency, achieving a zero-latency response. Resource comparison between Block RAM (BRAM) and distributed RAM, alongside extreme frequency stress testing, reveals that the IP core achieves a Worst Negative Slack (WNS) of 1.002 ns at 50 MHz, ensuring robust 1024 byte burst transmissions.
The importance of real-time semantic segmentation on edge devices is increasingly recognized in fields such as autonomous driving and intelligent robotics. While AI accelerators and lightweight models have significantly improved inference speeds, the unaccelerated pre- and post-processing stages have emerged as new system-level bottlenecks. To eliminate these bottlenecks, this paper proposes an end-to-end hardware acceleration system based on the Xilinx Zynq MPSoC. Designed via High-Level Synthesis (HLS), the proposed custom IPs efficiently execute computationally intensive tasks, including image resizing, normalization, argmax, and alpha blending. By leveraging a dataflow architecture, the pre-processing computation is entirely overlapped with the camera capture stage, effectively hiding the latency from the host CPU. Experimental results demonstrate that the pure hardware IPs achieve 13.09 times and 19.42 times speedups for pre- and post-processing, respectively. Consequently, the end-to-end throughput is improved by 1.86 times (achieving 68.44 FPS), and the total latency is reduced by 2.48 times compared to a CPU-based software baseline. Furthermore, the proposed system achieves an energy efficiency of 3.93 FPS/W, improving by 1.74 times over the CPU baseline, without compromising the segmentation accuracy yielding a mere 0.11 percent pixel discrepancy.
Visual odometry technology is gradually evolving towards deep semantic feature fusion, but the massive tensor computation overhead severely limits the deployment upper limit of the algorithm in power-constrained edge computing environments. This paper proposes a heterogeneous system FPGA hardware acceleration architecture for deep learning-enhanced visual SLAM. By designing an efficient multi-level on-chip hybrid cache and a high-concurrency pipelined systolic array, this paper effectively overcomes the memory bandwidth wall problem in the high-dimensional feature extraction and matching stages. Experimental data verification shows that, while maintaining the confidence accuracy of feature point output, the absolute trajectory error of this acceleration engine is stable at the level of 0.013 meters, its front-end integrated throughput reaches 65 FPS, and the overall dynamic power consumption is only ${6. 5}$ watts. This research constructs an embedded perception base that perfectly balances computing power, energy efficiency, and robustness, providing a powerful computing foundation for the field of autonomous driving in complex and unstructured environments.
Background/Objectives: Oral squamous cell carcinoma (OSCC) represents the most common oral malignancy, accounting for over 90% of cases, with a five-year survival rate of approximately 50%. Oral potentially malignant disorders (OPMDs) carry an inherent risk of malignant transformation to OSCC, yet early and accurate diagnosis of both conditions remains challenging in clinical practice, owing to lesion heterogeneity and inter-observer variability in histopathological interpretation. While existing reviews have addressed artificial intelligence (AI) models’ diagnostic performance in oral oncology, these have typically combined varied model types and have not restricted inclusion to studies based on a held-out test set. This scoping review maps convolutional neural network (CNN)-based models applied with explicit diagnostic intent primarily to OSCC, with exploratory coverage of OPMDs, across histopathological and clinical photographic imaging, limited to studies reporting performance on held-out test data separate from model development. Methods: Following PRISMA-ScR guidelines, a systematic search was conducted in August 2025 across Scopus, PubMed, and the Cochrane Library. Studies were included if they applied CNN-based models with explicit diagnostic intent for OPMDs and/or OSCC using histopathological or clinical photographic images and reported quantitative diagnostic performance on a held-out test set. Extracted data included model architecture, imaging modality, dataset characteristics, and performance metrics comprising accuracy, sensitivity, specificity, precision, recall, F1-score, and AUC (area under the receiver operating characteristic curve). Results: Thirty studies published between 2020 and 2025 were included. CNN-based models reported generally favorable diagnostic performance, particularly with histopathological images, with accuracy frequently exceeding 90%. Models utilizing clinical photographic images demonstrated greater variability, with accuracy ranging from approximately 76% to 97.5%, in the seven of 13 studies that reported it. Across both modalities, binary classification tasks generally outperformed multi-class differentiation. Inconsistent metric reporting, reliance on small single-institution datasets, and near-absent reporting of external validation were observed as recurring methodological limitations. Conclusions: CNN-based models reported high diagnostic performance for OSCC, particularly in histopathological binary classification settings, although evidence for OPMD subtypes remains limited. However, translating this into clinical practice will require investment in large shared data repositories, standardized reporting of performance metrics, rigorous benchmarking against human experts, and external and prospective clinical validation.
We study learning-augmented combinatorial optimization in the absence of externally supplied labels. We propose CLIO (Contrastive Learning with Iterative Optimization), a search-guided graph neural framework for the maximum independent set (MIS) problem. CLIO is designed around a closed-loop optimization paradigm: graph reduction simplifies the instance, a graph neural network (GNN) produces heuristic node scores, and a local deep search (LDS) procedure enforces feasibility and improves the candidate solution. Critically, the solution returned by LDS is converted into instance-specific pseudo-labels,which are then used to update the GNN through supervised and contrastive objectives. In this way, CLIO learns to solve an unlabeled MIS instance without requiring optimal solutions, external expert decisions, or reward engineering. The contribution of this work is therefore not a single neural architecture, but an integrated neuro-symbolic solving framework that enables label-free adaptation on a per-instance basis. Experiments compare CLIO with greedy and search heuristics, GNN-only and search-only variants, and time-limited solver references. In the ablation study on PPI graphs, the full framework achieved the highest average solution size among the methods except for the variant without graph reduction. LDS contributed the most to solution quality improvement; the closed loop contributed 0.1% and the contrastive loss contributed 0.4% to solution quality improvement, while graph reduction significantly reduced runtime. The current evidence demonstrates that search-generated pseudo-labels provide a viable learning signal for neural combinatorial optimization when no labeled data are available.
Electroencephalography (EEG)-based affective-state recognition may provide an objective neurophysiological complement to self-report scales and behavioral observations, but robust modeling remains difficult because EEG signals are non-stationary, temporally heterogeneous, and strongly subject-dependent. This study evaluated MFANet, a multi-scale feature fusion and neural architecture search-enhanced framework for binary classification of valence, arousal, and dominance. MFANet combines parallel one-dimensional convolutional branches with different receptive fields, channel-wise attention for adaptive feature recalibration, and a neural architecture search-guided long short-term memory module for temporal modeling. The framework was evaluated on the public DEAP and DREAMER datasets using five-fold, trial-grouped subject-dependent cross-validation in which all windows from one stimulus trial remained in the same outer fold and all data-dependent operations were confined to the corresponding training data. In DEAP ablation experiments, the complete model achieved mean accuracies of 97.20 ± 1.63% for valence, 97.32 ± 1.27% for arousal, and 97.14 ± 1.21% for dominance. Compared with the basic convolutional-recurrent baseline, multi-scale feature fusion and the searched recurrent module improved average accuracy by 10.17 and 10.87 percentage points, respectively, corresponding to relative gains of 12.5% and 13.4%. In participant-level paired comparisons based on the archived participant-level values, MFANet exceeded the dynamic graph convolutional neural network (DGCNN) by 4.13, 3.39, and 3.71 percentage points; bootstrap 95% confidence intervals were 3.29–5.15, 2.56–4.38, and 2.74–4.85, and all Holm-adjusted p-values were <0.001. A DEAP channel-count sensitivity analysis indicated that the full 32-channel montage changed accuracy by less than 0.4 percentage points relative to the prespecified 14-channel montage. These findings suggest that joint multi-scale representation learning and adaptive recurrent architecture selection can improve benchmark EEG affective-state recognition. Because the analysis used public healthy-volunteer datasets and self-reported affective labels, the framework should be regarded as a computational method for future clinical validation rather than a diagnostic tool.
Wind energy forecasting in complex terrain is a critical challenge for reliable power grid integration. This PRISMA 2020-guided structured review examines 134 eligible studies (2018–2026) that apply deep learning (DL) and other neural-network methods to wind speed forecasting, wind power prediction, and wind field downscaling, with particular emphasis on complex terrain. Four methodological pillars are analyzed: automated deep learning (AutoDL/NAS), multi-source data fusion, uncertainty quantification with emphasis on conformal prediction, and explainable AI (XAI). Only 24 studies (18%) explicitly evaluate mountainous, valley, coastal, or urban settings. CNN-based models are the most frequent architecture family, whereas AutoDL/NAS and automated hyperparameter optimization appear in only 11 studies, and no NAS framework has been evaluated in complex terrain. Reported fusion gains are study-specific (from about 13% RMSE reduction to over 50% MAE reduction) and are not directly comparable across datasets, metrics, and baselines. Conformal prediction offers finite-sample coverage guarantees under exchangeability, yet its empirical reliability under terrain-induced non-stationarity has not been evaluated. XAI appears in only five DL studies and is mostly applied post hoc. No study combines three or more pillars. We identify ten challenges and propose ten future directions toward integrated, terrain-aware, uncertainty-calibrated, and interpretable forecasting frameworks.
Gold volatility responds to shocks and sentiment, yet the two are rarely fused, and forecasts seldom become actionable. We link a multi-source fusion subsystem to a risk-warning layer through a Transformer whose three pre-registered hyperparameters are searched by particle swarm optimization (PSO-Transformer). It fuses video sentiment, headline scoring and market factors screened to 20 of 21 candidates; the decision side maps the predicted realized variance RVt/Mt (bp2, non-annualized) into three traffic-light tiers of a calibrated predictive distribution. On 5-min spot gold data (January–June 2025, 126 trading days), the PSO-Transformer attains a one-step MAE of 8.038, a 36.5% average improvement over seven baselines under protocol P1, where selection and evaluation share one window; under the leak-free protocol P2, the same pipeline attains 10.16 against 8.82 for the GRU, so the first place reflects the P1 design, not an independent out-of-sample comparison. Diebold–Mariano gains over GARCH, SVM and XGBoost survive a Holm correction. Walk-forward backtesting over 59 origins with persistence-anchored alerts yields 67.8% tier accuracy, 78.8% high-risk precision and 92.9% recall, with the neural engine fed into the same rule attaining 45.5% against 61.4%. One instrument, six months: a documented architecture, not a demonstrated transfer.
The monograph aims to provide a detailed analysis of current biometric authentication technologies using electrocardiographic (ECG) signals and voice as two fundamentally different but complementary modalities. The relevance of the work is due to the ongoing and evolving crisis of password-based security mechanisms (annual leaks affecting billions of user accounts, and generative speech synthesis technologies moving towards levels where the human voice can be imitated in a matter of minutes). The combination of ECG (an innate marker of life that is almost impossible to fake) with voice (which provides simple contactless verification at a distance) forms a logical and specialized premise for a new generation of biometric structures. The first part is devoted to the basic theory of biometric technologies: nine principles (uniqueness, stability, measurability, universality, resistance to counterfeiting, security, accuracy, standardization and acceptability) for building systems; a comparative analysis of nine physiological and behavioral modalities; nine functional modules of a typical biometric system; architectural solutions from autonomous to cloud-based and adaptive with Zero Trust support; a system of quality assessment metrics – FAR/FRR/EER/DET/minDCF curves – and approaches to ensuring demographic equity. The second part is devoted exclusively to ECG-based biometrics. The unique characteristics of the electrocardiographic signal, arising from anatomical, electrophysiological and neurohumoral factors, are further explained in physiological terms. All processing stages are described – from signal acquisition (with contact and non-contact methods), noise filtering (adaptive LMS filter or 1D U-Net based on neural networks), cardiac cycle segmentation using the Pan–Tompkins algorithm, autoencoder architectures, formation of morphological, temporal and spectral features for average heart beat analysis and dimensionality reduction based on PCA/LDA – up to classification and decision making. Comparison of two-class SVM, k-NN, decision trees and neural networks in different configurations was performed using four datasets (MIT-BIH, PTB, Fantasia, Lviv Biometric Data Set). The highest identification accuracy (98.90%, FAR 0.18·10⁻³, FRR 7.32·10⁻³) was achieved by a neural network with a hidden layer architecture (256, 128). It was experimented with deep embedding architectures (1D-ResNet, Transformer) as well as metric loss functions (Triplet Loss, ArcFace) and tested whether it is possible to deploy it on edge devices. The third section contains the full voice biometric chain: classical MFCC, LPC and PNCC descriptors, speech preprocessing methods (spectral subtraction, Wiener filtering, VAD and neural network optimization using Wave-U-Net), and finally four modern neural network architectures for speaker embedding-ECAPA-TDNN/WavLM/TitaNet/PyAnnote. In the framework of the experimental study, it is shown that TitaNet achieves the highest verification accuracy (0.96), and ECAPA-TDNN – 0.95 and scales better with increasing number of users; while WavLM, under difficult conditions, demonstrates the highest sensitivity to channel variations. The fourth section covers several practical implementation topics: international legal regulation (GDPR, AI Act, BIPA, ISO/IEC 24745 and 30107 standards), ethical aspects of the use of biometric data, architectural solutions for industrial implementation (design of a template database with approximate search algorithms: HNSW and IVF-PQ; cancelable biometric methods; fault tolerance; integration of BioAPI and FIDO2), implementation scenarios in the e-ID industry: banking, telecommunications, medicine, public administration and corporate security; the topic of the user interface - from ease of use to accessibility and economic analysis of costs against return on investment. Strategic development opportunities are identified: multimodal integration of ECG and voice at the level of feature fusion, scoring and decision-making; adaptive systems based on federated learning with differentiated privacy against threats from generative AI; transfer to decentralized identification based on Zero-Knowledge Proof. The scientific novelty of the work is determined by a new integrated concept of biometric authentication with a combination of ECG and voice; a comparative study of modern embedding architectures not only with an assessment of accuracy, but also with a determination of the scalability of models and resistance to spoofing attacks, based on the most modern generative models; as well as a quantitative measurement of how serious a threat these models pose to voice biometrics systems. The practical significance depends on the ability to directly apply the results to build biometric systems in finance, telecommunications, medicine and other areas of social activity. The monograph is designed for researchers, graduate students and students of specialties in information security, digital signal processing and machine learning, as well as for practicing specialists in the field of design and implementation of biometric authentication systems.
Early detection of leaks, bursts, process faults, and cyber-physical attacks is a direct lever for the sustainable management of water infrastructures: it reduces non-revenue water, the energy spent on treating and pumping lost water, and service interruptions (Sustainable Development Goal 6). Anomaly detectors for water distribution networks (WDNs) and wastewater treatment plants (WWTPs) are increasingly deployed through Artificial Intelligence of Things (AIoT) architectures, yet a detector that performs well on one system cannot be assumed to transfer to another with a different topology, sensor layout, demand pattern, and label availability. This systematic review examines the literature through this transfer problem, with graph neural networks (GNNs) as topology-aware models, tabular foundation models (TFMs; in the retrieved evidence, the TabPFN family) as candidate local screeners, federated learning as privacy-preserving adaptation, and digital twins and explainable artificial intelligence as validation and trust layers. It follows PRISMA 2020 and is registered retrospectively in the Open Science Framework (DOI 10.17605/OSF.IO/D7RXU). Peer-reviewed articles, theses, and preprints in English or Spanish were identified through modular block searches (Google Scholar, publisher sites, three Spanish repositories), author-supplied sources, and citation chasing, verified against primary records on 6–7 September 2026, screened in two stages, coded with a predefined framework, appraised with a six-item checklist, and synthesized descriptively because tasks, datasets, and metrics are not commensurable. The corpus comprises 150 unique works: 76 water-specific experimental studies, 21 benchmark or testbed studies, and 53 method, review, or contextual works. Graph-based approaches have the stronger water-specific evidence for topology-dependent tasks in WDNs, whereas TabPFN-family TFMs rest on general tabular-learning evidence and one water-adjacent prediction study, with no validated water anomaly-detection use. Only 6 of the 76 experimental studies evaluate an explicit transfer mechanism, federated evidence is limited to one water study, and no study reports an operator evaluation of explanations. The review contributes a transfer taxonomy with study counts, an evidence-class scheme, a gap map of demonstrated and proposed components on a transfer-aware AIoT layout, and an evaluation agenda based on source–target protocols, and identifies the evidence utilities need before transferable detectors can support sustainable operation at scale.
A clinical recommendation becomes care after someone confirms coverage, finds a facility that provides the service, arranges language support and transportation, books an appointment, and verifies that the visit took place. When no organization owns this work, it falls to the patient, and a missed service is recorded as nonadherence, even when the failure occurred within the health system. The review examines five adjacent fields, each addressing part of this problem, and develops the Care Plan Compiler, an agentic AI framework that treats the clinician-approved recommendation as a protected source specification and compiles it into a graph of tasks with named owners. A deterministic readiness checker releases a task once all prerequisites hold and classifies any unknown coverage, capacity, or accessibility as unresolved. Feasible plans are compared across a multidimensional burden vector, and a change in context triggers recompilation of the dependent tasks. A synthetic case of diabetic retinal screening illustrates the compilation, and a four-stage evaluation program sets a screening completed within the approved window and verified within 90 days of referral approval as its primary endpoint. The framework moves responsibility for missed care to the system layer where the failure occurs.
Multimodal Large Language Models (MLLMs) are emerging as a foundation for automated front-end engineering because they enable joint reasoning over screenshots and natural-language instructions. However, most front-end assistants treat this workflow as a set of isolated code-generation tasks, directly decoding framework-specific syntax for each request. This direct-decoding paradigm leaves three limitations unresolved: (i) it provides no transferable prediction target across tasks and frameworks, (ii) it offers no stable representation for localized edits and repairs, and (iii) it provides no auditable path from multimodal input to executable outcomes. In this paper, we propose DeltaUI, a unified framework that recasts generation, editing, and repair as predictions of transitions in a shared, framework-normalized UI state space, modeling UI change rather than directly decoding target code. DeltaUI begins by parsing heterogeneous front-end artifacts into a typed UI state graph that separates shared UI semantics from framework-specific surface syntax. It then uses a multimodal transition model to predict structured state operations by encoding the source state and multimodal evidence (e.g., screenshots, instructions, and issues), preserving explicit source-to-update correspondence. A deterministic framework-conditioned realizer converts the resulting state into target-framework code. Finally, an execution-aware critic drives a bounded refinement loop using compiler and render feedback, ensuring the full path remains auditable. Extensive experiments on DesignBench and Design2Code demonstrate that modeling UI-state transitions yields consistent gains over direct code-decoding baselines across multiple MLLM backbones while retaining effective transfer to real-world webpages. The code and implementation of DeltaUI are provided in the supplementary material.
Objective: Disease screening and diagnosis provide common issues for pathology and clinical laboratory medicine. These challenges include growing workloads, staffing shortages, diagnostic variability, and error susceptibility across the entire testing process. Although machine learning (ML) has been extensively used in image analysis and laboratory automation, studies that have already been written frequently focus on these areas independently, which prevents a comprehensive understanding of how ML affects diagnostic practice. This narrative review's goal is to compile the most recent data on machine learning applications in pathology and clinical laboratory medicine, with a focus on workflow efficiency, diagnostic performance, and the identification of laboratory errors throughout the entire testing process. Methods: Using predetermined search terms associated with machine learning, pathology, clinical laboratory medicine, and diagnostic quality, we performed a narrative review of studies indexed in PubMed, Scopus, Web of Science, and IEEE Xplore. Four domains diagnostic image analysis, biomarker and outcome prediction, workflow automation, and laboratory error detection are combined to create ML applications. Applying deep learning (DL) models to whole-slide images (WSIs) in pathology has produced results that are on par with pathologists in controlled environments. By utilizing multianalyte and temporal patterns, ML-based systems in clinical laboratory medicine improve automated verification, patient-based quality control (QC), and error identification, frequently exceeding conventional rule-based techniques. Prognostic modeling and decision support are further enhanced by new multimodal frameworks that incorporate laboratories, clinical, and histopathologic data. Conclusions: This study emphasizes the potential and constraints of machine learning (ML) in diagnostic medicine, stressing the necessity of interpretability, generalizability, and regulatory supervision to enable clinically significant, integrated implementation.
The rapid growth of malware targeting modern computing systems necessitates accurate and scalable detection and classification mechanisms capable of operating in real-world environments. In this work, we present a two-stage AI-based malware analysis framework that combines binary malware detection with fine-grained malware family classification using convolutional neural networks trained on byteplot grayscale representations of executable files. The system employs an EfficientNet-B1–based detector to distinguish between benign and malicious files, followed by a custom CNN classifier that assigns detected malware to one of 34 families. The framework is evaluated on multiple large-scale datasets, including Malimg, the Microsoft Malware Classification Challenge, and the MaleX dataset. Experimental results show strong performance, with the malware classifier achieving 96.64% accuracy and an F1-score of 96.63%, while the detector demonstrates robust performance across different decision thresholds, achieving an F1-score of 94.03% at a threshold of 0.40. These results indicate high class separability, effective detection–classification integration, and the adaptability of the proposed approach to varying security requirements, supporting its practical deployment in malware analysis pipelines. To further demonstrate its practical applicability, the proposed framework is integrated into a Continuous Integration/Continuous Deployment (CI/CD) pipeline using Jenkins, enabling automated malware scanning as part of the software delivery process. This integration enables automated malware scanning of compiled executable artifacts prior to deployment, allowing the proposed framework to operate as a pre-deployment security gate within modern CI/CD pipelines. Experimental validation confirms successful end-to-end execution, where analyzed files are correctly classified, demonstrating reliable system integration, scalability, minimal human intervention, and suitability for deployment in real-world DevSecOps environments.
Lifestyle, genetic, and environmental influences are implicated in metabolic disorders, an emerging global health concern, including diabetes mellitus, dyslipidaemia, obesity, and hyperuricemia. The success of disease treatment and the prevention of long-term complications depend on the early identification and ongoing monitoring of metabolic markers. Even though these are accurate, traditional approaches for diagnosis are invasive, labor-intensive, and not amenable to real-time monitoring. By enabling the rapid, sensitive, and minimally invasive detection of biomarkers such as glucose, insulin, lipids, and uric acid, recent developments in biosensor-integrated biomedical technology have revolutionized metabolic diagnostics. This review provides a comprehensive overview of the various forms of biosensors, including electrochemical, optical, piezoelectric, and nanobiosensors, and their classification, mechanisms, and detection principles. High-performance materials such as graphene, carbon nanotubes, and metallic nanoparticles have been incorporated to significantly enhance biosensor performance, and microfluidic lab-on-a-chip systems have facilitated the convenience of collecting non-invasive samples of interstitial fluid, perspiration, or saliva. To ensure the analytical precision and clinical reliability of biosensor responses, validation methods based on bioinformatics and High-Performance Liquid Chromatography (HPLC) are highlighted. In addition, continuous monitoring, predictive medicine, and tailored healthcare management have been enabled by the incorporation of artificial intelligence (AI), machine learning (ML), and the Internet of Things (IoT) in wearable and intelligent biomedical devices. Widespread clinical translation is hindered by challenges like sensor drift, biofouling, calibration variability, and regulatory constraints, although tremendous progress has been achieved. Future biosensor technology will be developed towards anticipatory, patient-focused, precision-based management of metabolic disorders by synergistic implementation of nanotechnology, AI-enabled analytics, and telemedicine.
Unsolicited and scam text messages are a compelling case for on-device filtering, but many recent detectors are based on deep or transformer models that may be large or slow enough to be challenging to run on phones. We evaluate eight classical classifiers—Multinomial and Complement Naive Bayes, Logistic Regression, a linear Support Vector Machine (SVM), a Decision Tree, a Random Forest, a small multi-layer perceptron and XGBoost—on the de-duplicated Short Message Service (SMS) Spam Collection (5,171 messages) from the University of California Irvine (UCI) repository using term frequency-inverse document frequency (TF-IDF) unigram and bigram features, tuned inside a nested ten-fold cross validation, comparing predictive quality, serialized model size and single-message inference latency. The linear SVM, Logistic Regression and Multinomial Naive Bayes form a top tier, with cross validated spam-class F1 scores of 0.951, 0.951 and 0.950 respectively—with no pairwise statistically significant differences among them—occupying 223–340 KiB and classifying a message in about 1 ms via a Python pipeline running on a desktop central processing unit (CPU). The Random Forest and the neural network made no accuracy improvements, while being roughly 13 and 18 times larger than the linear models. A study of vocabulary size shows that constraining the linear SVM to 2,000 terms reduces the model size to 88 KiB at a negligible sacrifice of 0.007 F1. A library-free export of Multinomial Naive Bayes, Logistic Regression and the linear SVM matched the scikit-learn predictions on all 1,035 held-out messages for possible Android implementation; on-device timings remain future work. Error analysis reveals that remaining errors tend to cluster in very short, conversational or announcement-like messages.
Improper segregation of waste at the source leads to inefficient recycling, environmental pollution, and health hazards. This paper presents an Arduino and ESP32-CAM based smart waste segregation system that combines sensor-based detection with machine learning-based image classification. The ESP32-CAM captures images of waste items, while a trained model identifies categories such as biodegradable, medical, plastic, metal, glass, and wood. Moisture, inductive, capacitive, and proximity sensors provide additional material information. Arduino processes the sensor readings and classification output and controls servo motors to direct waste into the appropriate bins. The proposed hybrid approach aims to reduce manual effort, improve sorting reliability, and support cleaner waste management. The system can be extended with IoT-based monitoring for use in homes, hospitals, institutions, and smart city environments.
The lightweight design and reliability of physical unclonable functions (PUFs) are critical for IoT security. In this work, we propose a strong PUF based on spin-transfer torque magnetoresistive random access memory (STT-MRAM). For the first time, we adopt a logic-in-memory (LIM) structure known as Computational Random Access Memory (CRAM) in order to promote massive parallelism and low area overhead. In our simulations of the results, the proposed PUF macro occupies 2960.4 μm2 in total, 5.77 μm2 per cell, and presents a high bit rate of 1.42 Gb/s. In addition, we introduce the TRIM circuit and algorithm into the response generation process and develop two TRIM algorithms to optimize its reliability. The proposed PUF presents an excellent average intra-chip Hamming distance of 0.288% under the temperature range from −10 °C to 150 °C. The PUF also achieves an inter-chip Hamming distance of 50.003% and generates 22N/3 challenge–response pairs (CRPs), which surpass other state-of-the-art MRAM strong PUFs. More importantly, the proposed PUF shows excellent machine-learning resistance. This work provides a new strong PUF solution for resource-limited system-on-chip (SoC) chips for edge applications.
EEG emotion recognition typically follows the standard signal-to-label paradigm. It maps EEG signals to discrete labels via feature engineering and machine learning, yet neglects the rich semantic information inherent in the task. This purely data-driven approach easily overfits to subject-specific physiological patterns, severely degrading its generalization on new subjects. Here, we propose a Semantically Guided Domain Adaptation (SGDA) method for cross-subject EEG emotion recognition. Specifically, SGDA projects non-stationary EEG features into an emotional semantic space, leveraging the discriminative power of structured linguistic embeddings to enhance class separability. To construct this unified space, a pretrained language model is utilized to extract linguistic embeddings from emotion text labels, which serve as invariant semantic prototypes to guide cross-domain feature alignment. Moreover, to facilitate robust inference, an adaptive distribution-aware feature fusion strategy dynamically aggregates multi-source knowledge based on the distances between target samples and source centroids. Extensive experiments on five benchmarks demonstrate that SGDA achieves state-of-the-art performance in cross-subject scenarios. The source code is publicly available at https://github.com/EDH777/SGDA.
Faults are the principal impediment to achieving elevated levels of efficiency, reliability, and longevity in photovoltaic (PV) systems. In this article, we reviewed solar PV fault diagnostics techniques and compared different fault signature interpretations based on three primary data sources: electrical measurements, environmental conditions, and thermal inspection images. Five overlapping diagnostic perspectives are discussed: measurement modalities (electrical and thermal), statistical methods, optimization-based methods, machine learning techniques. The different approaches are compared regarding: the extent to which faults can be detected; quality of the diagnostic evidence that is generated; what measurements are required for diagnostics; how the topology of the system affects diagnostics; ability to localize faults; under what conditions a diagnosis is validated; and at what time interval the data needs to be acquired. Overall, our evaluation shows there is no single optimal method for identifying faults of all types, primarily because of differences in datasets, definitions of faults, validation conditions and timing measures which make it difficult to make direct comparisons. Finally, we identify areas for future research; Hybrid Physical/Data-Driven Models, Digital Twins, Embedded Inference, Standardized Datasets, Learning with Limited Labels
Aim: To appraise the documented performance and limitations of a masked/unmasked facial-recognition prototype and to specify a testable deep-learning architecture for subsequent evaluation.Design: A technical appraisal of thesis-reported prototype outcomes and a proposed but unimplemented architecture; no new comparative experiment is reported.Methods: The historical system combined an OpenCV deep-learning face detector with support vector machine (SVM)-based identity classification. The proposed system combines landmark-assisted face detection and alignment, ArcFace-style embeddings, cosine-similarity matching, an unknown-identity decision and optional ONNX Runtime inference. Evaluation requirements are defined for identity partitioning, mask conditions, multiple-person scenes and error reporting.Results: The thesis account attributes a face-detection accuracy value of 99% and approximately 80% live identity-classification accuracy to the prototype, while also presenting values of 97% for masked faces and 87% for unmasked faces elsewhere. Underlying test counts and predictions are unavailable, and these figures cannot be reconciled or recalculated. No accuracy or latency findings are available for the proposed architecture.Conclusion: A detector–embedding–matching workflow offers a technically justified route for further study, but its effectiveness, fairness, operating speed and suitability for deployment remain unestablished without reproducible evaluation.
Physical Unclonable Functions (PUFs) generate unique chip identifiers from manufacturing variations, but transmissions over public networks can be exploited through eavesdropping. To improve variation resilience, the proposed SRPFD-PUF employs bit self-test (BST) for stable-CRP selection, while response-bit tolerance is considered in an illustrative authentication flow. By minimizing additional circuit overhead, strong baseline PUF metrics are maintained at low hardware cost, supporting integration in Internet of Things (IoT) devices. An illustrative PUF-based authentication flow is presented to demonstrate a potential IoT application; end-to-end protocol implementation and security validation remain outside the scope of this study. The three-condition BST screen yields a minimum simulated logarithmic reliability score of 98.93% under the evaluated PVT conditions within 0.675–1.05 V and −40–150 °C. With the circuit architecture and LFSR configuration disclosed, the evaluated machine-learning attacks yield prediction accuracies of 47.68–52.46%, indicating resistance to the tested modeling methods. The TSMC 16 nm FinFET physical design occupies 0.0117 mm2, while post-layout simulation estimates 0.3089 mW at 100 MHz; no fabricated-silicon measurements are reported.
As novel hardware architectures are explored for the purposes of machine learning, including neuromorphic chips, asynchronous systems, and ASICs, effective learning algorithms are highly desirable. While backpropagation is the standard for differentiation on standard hardware, its requirement for global synchrony and exact differentiability limits the exploration of exotic algorithms on non-traditional substrates and devices. Zeroth-order (ZO) optimisation methods, such as SPSA and weight perturbation, offer a compelling alternative by estimating gradients through inference-only passes; however, these methods have historically suffered due to the `curse of dimensionality', where random perturbations become increasingly misaligned with the true gradient in high-dimensional spaces. In this work, we move beyond blind random noise by deriving zeroth-order search directions from activation geometry available during the forward pass. This leads to a family of Spectral Zeroth-Order methods, SZO-$\kappa$, whose directions solve local response-based gradient-guessing objectives under explicit stochastic assumptions. These methods explore feature-supported directions that are more likely to produce informative loss responses. Using multilayer and convolutional networks, we analyse gradient alignment, bias-variance structure, and training stability of these methods relative to standard perturbation baselines and backpropagation.
Post-harvest preservation of fruits and vegetables is a major challenge because quality changes rapidly when temperature, relative humidity, ventilation and ripening gases are not managed appropriately. Conventional storage often relies on periodic manual inspection, which can miss short-term environmental changes and delay corrective action. This survey reviews IoT-enabled monitoring and intelligent storage approaches for fresh produce, with emphasis on low-cost embedded systems suitable for small farms, shops and local storage facilities. The reviewed literature covers wireless monitoring of ambient parameters, cloud-connected cold-storage notification, machine-learning-based freshness classification, and fruit-specific temperature and humidity monitoring. Their sensing strategies, communication methods, control capabilities and validation limitations are compared. The review identifies gaps in crop-specific thresholds, sensor calibration, automatic control, power reliability, connectivity resilience and long-duration field validation. Based on these findings, a proposed ENTC project architecture is outlined using an ESP32, temperature-humidity sensing, a gas sensor, a display, a cloud dashboard and relay-driven ventilation or cooling. The paper presents a practical evaluation plan; it does not claim experimental results for the proposed prototype.