2026-11-01

10 repos · 0 papers · 61 scholar

Showing 71 of 71 items

GitHubsynthesis-pnr

awesome-data-analysis

🚀 500+ curated resources for Data Analysis & Data Science: Python, SQL, Statistics, ML, AI, Visualization, Cheatsheets, Roadmaps, Interview Prep. For beginners and experts.

2,157320
GitHuboptimization

betterplayer

Production-grade video player for Flutter (Android, iOS, Web, macOS, Windows, Linux) - HLS, DASH, DRM (Widevine & FairPlay), caching, PiP, VAST ads, and vertical video feeds.

Dart1,0501,400
GitHuboptimization

VRKA-Android

Native Android media downloader for VRKA with direct extraction and browser-assisted stream discovery.

Kotlin20322
GitHubmodel-compression

dsh-sieve

dsh-sieve: context engineering & token optimization plugin for DeepSeek Harness (DSH) — tool output filtering, context pruning, progressive skill disclosure. 36% smaller payload in offline replay. DSH 上下文管理与 token 优化节省插件。

TypeScript1094
GitHubmodel-compression

OTel

Open Telco (OTel) AI: an open ecosystem of telecom models, datasets, benchmarks, and reproducible training and inference recipes. A shared AI foundation that gives operators the freedom to choose their models, software, hardware, and infrastructure.

Python359
GitHubmodel-compression

edgejev

离线可用的本地类型化决策:4 核 CPU 单题 15.6ms。Local & offline Jev / System One inference on CPU — ONNX + INT8, no torch at runtime. 支持 laya / kev / PlayJev

Python344
GitHubsynthesis-pnr

heypcb-mcp

HeyPCB for AI agents: an MCP server that gives Claude Code, Codex, Cursor and claude.ai a PCB engineer. npx -y @heypcb/mcp

TypeScript335
GitHubai-hardware

HeliBit

A Neuromorphic Computing Architecture Based on Parametric Helical Trajectories and Topological 64/256-Bitboard Representations for Extreme Energy- Efficient Language Models

Python00
Scholarhardware-design

A 28-nm, 8.3-mm2, 76.9-μJ/Pairing Compiler-Driven Crypto-Processor for Hybrid ZKP--PBC Application With Software--Hardware Co-Design

Zero-knowledge proof (ZKP), one of the most popular privacy-preserving schemes, enables the prover to convince the verifier of a certain statement’s correctness without leaking any private information. Among them, the Pairing-based succinct non-interactive argument of knowledge (zkSNARK), like Groth16 and Plonk, featuring succinct proof size and constant-time verification, is deployed across a promising ecosystem. However, the heavy operations of proof generation that grow rapidly with application sizes and the frequent verification that requires costly Pairing severely hinder broader adoption. Moreover, there are many other cryptographic schemes also constructed based on Pairing, such as Boneh–Lynn–Shacham (BLS) signature, functional encryption (FE), identity-based encryption (IBE), and so on. Unfortunately, existing synthesis-based ZKP accelerators suffer from low area efficiency and incomplete operator coverage, causing poor overall speedup and expensive deployment cost. This work presents the first silicon-proven crypto-processor that supports complete proof generation/verification phases and can adapt to other Pairing-based cryptography (PBC). To achieve high flexibility while maintaining cost efficiency, we employ holistic optimizations across algorithm, architecture, and compiler levels, including the hybrid-grained instruction set, dedicated memory system, reconfigurable datapath, custom Pairing compiler, and utilization-oriented timing scheduling. Fabricated in 28-nm CMOS, it covers all related operators and achieves a <inline-formula> <tex-math notation="LaTeX">$6.4\times $ </tex-math></inline-formula> area reduction compared to prior synthesis-based work LegoZK, which currently makes it the most cost-effective candidate for the practical deployment of ZKP–PBC applications. As a case study, the processor performs <inline-formula> <tex-math notation="LaTeX">$2^{16}$ </tex-math></inline-formula>-gate proof generation/verification with 0.15 J/0.11 mJ, offering two orders of magnitude energy savings over a 14-core 2.4-GHz CPU. For the Pairing operator, it achieves development agility to enable fast design space exploration. It delivers <inline-formula> <tex-math notation="LaTeX">$17\times $ </tex-math></inline-formula>/<inline-formula> <tex-math notation="LaTeX">$1.8\times $ </tex-math></inline-formula> improvements of area-time product (ATP)/energy compared to the prior ASIC work.

Xiang-Ren Chen +80IEEE Journal of Solid-State Circuits
Scholarhardware-design

Search-Guided Reasoning Trace Curation for Verilog Generation

Recent advances in large language models (LLMs) have improved Verilog generation. However, most training pipelines distill natural-language reasoning traces whose intermediate design decisions are weakly structured and are primarily selected according to the correctness of the final generated RTL. We present MCTS-SFT, a verifier-guided framework for curating reasoning traces for supervised fine-tuning (SFT) of Verilog generation models. MCTS-SFT formulates trace construction as search over intermediate reasoning states and uses Monte Carlo Tree Search (MCTS) to propagate functional feedback from complete Verilog rollouts to earlier design decisions. Within this framework, we introduce Plan-MCTS, which searches over compact hardware-design plans that explicitly represent interfaces, timing, state, behavior, and update rules. We further study plan-preserving supervision, in which the searched structured state is retained as part of the downstream SFT target. Our experiments find that Plan-MCTS achieves the highest functional trace yield, with 84.2% correct traces compared with 81.9% for one-shot structured planning and 82.0% for unguided tree exploration. For identical Plan-MCTS trajectories and Verilog outputs, retaining the explicit searched plan during SFT raises pass@10 from 52.56% to 57.7%, demonstrating that the representation of intermediate design decisions materially affects downstream learning. In a separate 6,000-trace study, MCTS-curated supervision improves pass@10 over full-response chain-of-thought distillation by up to 3.7 points while reducing downstream generation length by approximately 50% on VerilogEval-v2. These results demonstrate that structured hardware planning, verifier-guided search, and explicit intermediate-state supervision provide complementary mechanisms for constructing more effective and token-efficient training data for Verilog generation.

Matthew DeLorenzo +60
Scholarhardware-design

A Resource-Efficient and Precision-Scalable FPGA Accelerator for Spiking Neural Networks Using Minifloat Arithmetic

Spiking neural networks (SNNs) are well suited to event-driven edge intelligence, but the arithmetic complexity, memory footprint, and power consumption associated with conventional floating-point implementations can limit their deployment on resource-constrained hardware. This paper presents a precision-scalable and resource-efficient SNN hardware accelerator architecture based on parameterized minifloat arithmetic. The proposed architecture implements recurrent leaky integrate-and-fire (LIF) neurons and leaky-integrator (LI) readout neurons using a unified minifloat datapath whose word length and mantissa width can be configured at synthesis time. A valid-driven arithmetic subsystem employs pipelined minifloat multiplication and addition, while time-multiplexed processing enables arithmetic resources to be shared across synaptic accumulation, neuron-state updates, recurrent connections, and output processing. This organization reduces replicated arithmetic hardware and provides a scalable design approach for different SNN sizes and precision requirements. The architecture supports E8M23, E5M10, E5M6, and E4M3 formats within the same RTL framework, allowing precision, memory capacity, timing, and power to be traded without modifying the underlying neuron architecture. FPGA implementation results demonstrate that reducing minifloat precision substantially decreases arithmetic-resource utilization and estimated dynamic power, while also reducing the theoretical storage requirement of synaptic parameters. A recurrent event-based classification network is used as a hardware validation case, demonstrating that substantial precision reduction can be achieved while retaining competitive inference accuracy. The results establish parameterized minifloat computation as an effective basis for scalable, low-resource, and energy-efficient FPGA SNN accelerators.

Ali Mehrabi +10
Scholarsynthesis-pnr

Ultra-Broadband and Wide-Angle Solar Absorber With Double Annular Design Based on TiN

Titanium nitride (TiN) has attracted extensive attention in photothermal conversion and photovoltaic fields due to its remarkable plasmonic properties and excellent high-temperature thermal stability. In this work, we propose an ultra-broadband solar absorber operating in the 300–2500 nm wavelength range, which adopts a classical metal-insulator-metal (MIM) configuration consisting of a top nanostructured TiN metasurface, a SiO2 dielectric spacer layer, and a continuous TiN reflective substrate. Different from complex multi-layer or multi-size resonator designs, the proposed absorber achieves broadband perfect absorption through a single-layer double-annular structure that excites multiple coupled resonant modes, which greatly simplifies the fabrication process. Numerical results calculated by the finite-difference time-domain (FDTD) method demonstrate that the average absorption of the proposed absorber reaches 94.17% in the 300–2500 nm solar radiation band. Benefiting from the high rotational symmetry of the annular structure, the device exhibits polarization-insensitive absorption behavior and maintains high absorption efficiency under wide incident angles, with average absorbance exceeding 90% even at an incident angle of 50° for both TE and TM polarizations. Physical mechanism analysis reveals that the ultra-broadband absorption performance originates from the synergistic coupling of surface plasmon resonance (SPR), magnetic resonance (MR), and Fabry-Perot cavity resonance (CR). Furthermore, the calculated solar energy harvesting efficiency under the standard AM 1.5G spectrum reaches as high as 93.57%. With the advantages of simple geometry, good fabrication compatibility, and robust optical performance, the proposed absorber provides potential research reference for future high–efficiency solar energy harvesting and photothermal conversion systems.

Feng Zhang +30IEEE Photonics Journal
Scholarsynthesis-pnr

OCHRA: Integrated Design of a Lightweight, High-Payload Compliant Humanoid Arm

Humanoid arms must combine human-compatible geometry, payload capacity, and physical compliance despite competing structural and actuator demands. This paper presents the Optimized Compliant Humanoid Robotic Arm (OCHRA), an integrated framework coupling configuration search, lightweight skeletal design, a compact parallel-axis wrist, and series elastic actuation. Candidate configurations are evaluated using workspace, dynamic joint-torque demand, and orientation dexterity, followed by multi-load-case topology optimization and actuator allocation. The resulting 12.45 kg, 7-DoF prototype sustained a 15 kg static flange load, corresponding to an arm-only payload-to-weight ratio of 1.205. It also executed an outbound-and-return motion with the same load without exceeding continuous torque ratings. Intrinsic joint-torque sensing captured directional changes in estimated end-effector wrenches and distinguished nominal 0.2 N quasi-static load increments at three fixed postures without a wrist force/torque sensor. Together, these results demonstrate integrated payload, compliance, and intrinsic sensing functions under the reported test conditions.

Xiu-Bo Xia +40IEEE Robotics and Automation Letters
Scholaraccelerators

A Hybrid Inner--Outer-Product SRAM Compute-in-Memory GEMM Engine With Precision-Adaptive Systolic Dataflow

This article presents an efficient SRAM compute-in-memory (CIM) macro design for general matrix multiplication (GEMM). A hybrid inner–outer product (HIOP) architecture is designed to enable long-time weight reusing and short-time input reusing, reducing 1.57–<inline-formula> <tex-math notation="LaTeX">$2.69\times $ </tex-math></inline-formula> overall SRAM access. A data-precision-adaptive dataflow is developed to scale the CIM macros with balanced bandwidth (BW) across varying data precisions. Our CIM-friendly dataflow obtains 19%–38.5% reduction in operation cycles with the CIM capability of concurrent read/write and computing. Fabricated in 22-nm technology, the CIM macro achieves energy efficiency (EE) of 56.2 TOPS/W for GEMM and 71.65 TOPS/W for multiply–accumulate operation (MAC) in 8-bit inputs and weights. The macro and system area efficiency (AE) are 3.18 and 1.89 TOPS/mm<sup>2</sup>, respectively, which is <inline-formula> <tex-math notation="LaTeX">$1.42\times $ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$1.12\times $ </tex-math></inline-formula> compared with prior state-of-the-art (SOTA) designs.

Wen-Jie Ren +90IEEE Journal of Solid-State Circuits
Scholaraccelerators

A Mixed-Signal SAT Solver for Exploring Nonideal Effects and Scalability in Analog Computation

Recent years have seen growing interest in custom and nature-inspired computing methods for solving computationally challenging combinatorial optimization problems. The Boolean satisfiability (SAT), a canonical nondeterministic polynomial (NP)-complete problem, has remained a focal point due to its wide applicability in artificial intelligence, cryptography, and formal verification. This article presents a custom mixed-signal hardware SAT solver that maps the cost function of three-SAT instances onto an array of current-coupled computational nodes. The system progressively minimizes the cost function while employing an integrated perturbation mechanism to escape local minima. The proposed architecture is implemented in a standard 65-nm CMOS technology and experimentally evaluated. A central open question in the analog computation domain is the scalability of analog and mixed-signal architectures under inherent nonidealities. Beyond demonstrating functional correctness, this work systematically characterizes the impact of deviations inherent to analog computation, including coupling strength mismatch and nodal threshold variations. Measurement results reveal that solver performance is highly sensitive to random mismatch and largely insensitive to threshold variation. These effects are further elucidated through controlled experiments, such as random clause shuffling, selective sample pruning, and autozeroing, offering design guidance for scalable analog and mixed-signal optimization accelerators. Experimental results show that the fabricated integrated circuit achieves a median solution time of $9.75~\mu $ s per instance with a median energy consumption of 118 nJ across the uniform random 50-variable SATLIB benchmark instances.

Ahmet Yusuf Salim +40IEEE Journal of Solid-State Circuits
Scholarmodel-compression

Surrogate assisted artificial neural network and NSGA II framework for continuous multi-objective optimization of sheet metal blanking

Sheet metal blanking optimization is traditionally performed using design-of-experiments (DOE) techniques that evaluate only discrete parameter combinations, limiting the identification of optimal operating conditions within the continuous design space. This study proposes a surrogate-assisted Artificial Neural Network (ANN) and NSGA-II framework for continuous multi-objective optimization of sheet metal blanking. A unified ANN surrogateab model was developed using 100 experimental observations obtained from Taguchi L25 orthogonal array experiments on four engineering materials (SS316 steel, C110 copper, AA1100 aluminium, and C26000 brass). The model predicts burr height from sheet thickness, punch–die clearance, and material type, while analytical formulations were integrated to evaluate blanking force, shearing energy, and specific work, resulting in a four-objective optimization framework. A systematic architecture search involving 420 ANN models identified the optimal 6-12-1 network, achieving an MSE of 6.71 × 10−6, RMSE of 0.00259 mm, and R2 of 0.995. Model robustness was further verified using repeated five-fold cross-validation, sensitivity analysis, and comparative evaluation against six alternative surrogate learning models. The trained ANN was coupled with NSGA-II to generate 140 continuous Pareto-optimal solutions beyond the discrete DOE grid, enabling identification of balanced knee-point solutions through normalized Euclidean distance from the utopia point. Experimental validation of twelve representative Pareto solutions demonstrated excellent agreement between predicted and measured burr heights, with a mean absolute error of 0.0023 mm and a mean absolute percentage error of 3.31%. The proposed framework provides an accurate and computationally efficient decision-support methodology for continuous multi-objective optimization of sheet metal blanking and establishes a surrogate-assisted optimization strategy that can be extended to intelligent manufacturing optimization problems following appropriate retraining and validation.

S. Bhoir +10Discover Mechanical Engineering
Scholarmodel-compression

An Explainable Clinical Decision Support Framework Using Adaptive Marine Predators Algorithm-Optimized Improved CNN For Accurate Weighted Fuzzy Production Rule Extraction

Clinical Decision Support Systems (CDSSs) need a predictive accuracy as well as interpretability to assist clinicians in disease diagnosis and risk assessment. Traditional Convolutional Neural Networks (CNNs) manage to capture complex nonlinear relations but suffer from opacity in terms of outcomes, are sensitive to initialization of hyperparameters and cannot be utilised for explainable decision-making in healthcare. This research presents a novel explainable as well as hybrid framework to integrate an Adaptive Marine Predators Algorithm (AMPA)-optimized Improved CNN with the extraction of Weighted Fuzzy Production Rule (WFPR) to bridge predictive accuracy and interpretability. The upgraded CNN employs convolutional filters operating at different scales, as well as residual connections that allow for the passage of gradients, which address the problem of vanishing gradients. In addition, it uses attention to weight features that are clinically relevant. The Adaptive Marine Predators Algorithm is put into play in the weight and hyperparameter optimization of the CNN. The model uses adaptive Brownian and Levy motion to ensure global convergence and to avoid local optima or loss in high-dimensional search spaces. The post Adaptive Marine Predators Algorithm EM Asymmetrical Black Hole Optimization appeared first on Javatpoint. The outputs of CNNs are subjected to fuzzy membership modelling. Therefore, interpretable WFPRs are inferred, transforming underlying complex learned representations into easily understandable decision rules by clinicians. This framework was tested on various clinical datasets which showed increased classification accuracy, strong training stability as well as better rule interpretability. Overall, the CDSS architecture is transparent and scalable and can ensure reliable health care insight.

K. Sangeetha +10
Scholarmodel-compression

A neural network surrogate model-based time-domain sparse Bayesian learning method for CA mortar void identification of slab track system

Concealed voids of the cement-emulsified asphalt (CA) mortar layer in the slab track system pose considerable structural risks to high-speed railways, necessitating early detection. Finite element model updating methods offer global structural health assessment but are computationally demanding due to numerous simulations in high-dimensional optimization. To overcome these limitations, a novel damage identification method integrating a parameterized neural network surrogate model with time-domain sparse Bayesian learning is proposed. The surrogate model combines convolutional neural networks and long short-term memory networks, employing a dual-channel architecture for static structural parameters and dynamic impacts, further enhanced by position encoding and residual learning to predict acceleration sequences. Parameter optimization in this framework is performed with an improved particle swarm optimization algorithm with a local search strategy. The surrogate model achieved predictions with an average mean squared error of 0.0022 and an R -squared value of 0.913. The feasibility of the proposed method was validated on a scaled model of the slab track system. The numerical and experimental results demonstrate that the proposed method can successfully identify the location and severity of the CA mortar void and quantify the associated uncertainties. Furthermore, the proposed method significantly enhances computational efficiency, which offers a viable technical foundation for structural health monitoring of the slab track system.

Bi-Wei Zhang +30Structural Health Monitoring
Scholaroptimization

Global Scheduling of Transient Parameter Materialization for Sharded Training

Parameter-sharded training reduces device-memory use by storing model parameters as shards and reconstructing full parameters only before the computations that consume them. This materialization is implemented through frequent all-gathers, which can become exposed waits in each forward–backward training iteration. Hiding these waits is not only a communication-volume problem. It requires deciding where to launch each all-gather and how long the materialized full-parameter buffer should remain live. All-gather prefetching can overlap communication with earlier computation but extends buffer lifetime. Retaining a buffer after its forward-pass use can eliminate a re-gather before backward-pass reuse, but consumes transient memory headroom needed by other materialization decisions. We present Chorus, a compiler-runtime framework for transient parameter materialization in memory-constrained sharded training. Chorus addresses a limitation of local runtime policies that make decisions at hooks, module boundaries, or configuration thresholds. They can issue individual prefetches or suppress individual releases, but they do not jointly compare future all-gather opportunities, buffer lifetimes, and memory-headroom conflicts across the iteration. Chorus lowers the forward and backward passes into a whole-iteration graph that makes all-gathers, use-site waits, and releases explicit. It then builds a profile-guided block model and solves a mixed-integer scheduling problem that jointly chooses all-gather launch positions and iteration-local retention decisions under time-varying memory headroom. The selected schedule is realized by graph rewriting, which moves communication while preserving use-site waits, fuses compatible prefetch operations, and removes redundant releases and re-gathers. Cross-iteration persistent retention is finalized during warmup using measured runtime memory behavior. We implement Chorus on top of torch.compile and evaluate it on 7B to 13B dense Transformer models and one MoE model across 4 to 16 A100 GPUs. In the ZeRO-style stack, Chorus improves throughput by 1.10× over DeepCompile across the primary workloads. In the FSDP-style stack, the Chorus implementation on SimpleFSDP improves throughput by 1.14× over SimpleFSDP across the corresponding dense-model workloads. The benefits are greatest in network-constrained settings, where all-gather stalls are more exposed. Chorus stays within the configured memory budget and achieves a 1.35× compilation-time speedup over DeepCompile.

Zuo-Yan Zhang +30ACM Transactions on Architecture and Code Optimization (TACO)
Scholarframeworks

EcoMLC: An Energy-Efficient Machine Learning Compilation Framework via Cache Reuse and Frequency Scaling

While modern machine learning compilers (MLCs) are pushing deep neural network (DNN) inference latency to its theoretical limits, the accompanying energy overhead has become a critical bottleneck for large-scale deployment. Excessive energy consumption not only increases electricity costs, but also affects the sustainability of edge platforms. Traditional performance-first optimizations often lead to significant energy waste, as they overlook the complex interplay between kernel configurations, cache behaviors and frequency scaling. This article presents EcoMLC, an energy-efficient machine learning compilation framework. EcoMLC introduces a reverse scheduling policy combined with an explicit cache priority mechanism, mitigating cache thrashing inherent in standard least recently used (LRU) policies. EcoMLC employs an end-to-end frequency-sweeping analysis to identify optimal energy operating points and enables fine-grained tradeoffs between performance and energy under diverse service level objective (SLO) constraints. Experimental results demonstrate that EcoMLC effectively optimizes various workloads, achieving an average of 43% and 26% energy savings (in J/request and J/token) over NVIDIA’s TensorRT and vLLM, and 22% energy savings on an AMD GPU compared with vLLM. For certain large language model (LLM) configurations, EcoMLC achieves a 33.9% reduction in energy consumption at the cost of a marginal 7% decline in performance.

Chao-Yang Gu +60ACM Transactions on Architecture and Code Optimization (TACO)
Scholarframeworks

NeuroComp: A Cross-Platform Compiler for Efficient SNN Deployment on Neuromorphic Hardware

Spiking Neural Networks (SNNs) offer substantial energy efficiency advantages, yet deploying them across diverse neuromorphic hardware remains challenging. Existing toolchains exhibit a critical tradeoff: platform-specific frameworks lack portability, while low-level programming interfaces demand extensive hardware expertise and produce non-portable code. Furthermore, conventional deep learning compilers fail to handle the unique temporal dynamics of SNNs and the idiosyncratic constraints of neuromorphic architectures. To bridge this gap, we present NeuroComp, a cross-platform compiler that automatically generates functionally correct and high-performance code for heterogeneous neuromorphic backends. Our core insight is that the diverse execution behaviors of neuromorphic hardware can be universally modeled as parallelizable scalar computations expressible via nested loops. Building on this, NeuroComp introduces a unified multi-level loop abstraction to systematically align SNN computational semantics with hardware capacities. To efficiently navigate the vast implementation space, we propose a decoupled two-stage scheduling strategy. The first stage performs batch partitioning to guarantee functional correctness and strict architectural compliance, while the second stage optimizes inter-batch execution orders for maximum data reuse, guided by lightweight, hardware-specific cost models and heuristic pruning. We evaluate NeuroComp on three representative architectural paradigms: an instruction-extended CPU, a dedicated accelerator, and a many-core hardware. Experimental results demonstrate that NeuroComp consistently achieves competitive performance compared to meticulously hand-optimized implementations and significantly outperforms existing generic automated compilers, effectively minimizing development effort while unlocking cross-platform portability.

Chao Xiao +50ACM Transactions on Architecture and Code Optimization (TACO)
Scholarai-hardware

Kramers–Kronig Receiver Enhanced by Photonic Reservoir Computing

The rapid growth in bandwidth demand in modern optical networks has driven the need for high-data-rate modulation techniques such as 16-QAM. However, these methods typically depend on complex and power-intensive coherent receivers. To address this, the Kramers–Kronig (KK) algorithm enables simplified demodulation using a single photodiode. This work proposes a hybrid receiver combining an all-optical Photonic Reservoir Computing (PRC) system with a KK direct-detection receiver to improve performance. The KK algorithm reconstructs phase information digitally, while PRC acts as an optical equalizer to mitigate dispersion and nonlinear effects over a 70 km fiber link, ensuring accurate phase recovery. Performance comparisons using various Support Vector Regression (SVR) kernels like Linear Kernel, Radial Basis Function (RBF) Kernel, Anova RBF Kernel, Arc Cosine Kernel, Regression models like Linear, Ridge, and Lasso regression, Ensemble Based Learning (Boosting) model and Classification model (Naive Bayes) at the reservoir’s readout layer was performed. It shows that the Naive Bayes classifier achieves the better performance in terms of BER and Q factor. Overall, results confirm that integrating PRC with the KK receiver significantly reduces BER compared to conventional KK detection, evaluated using BER, Q-factor, and constellation diagrams.

Bhubneswar Gosain +30IEEE Photonics Technology Letters
Scholarai-hardware

FrozenSFL: Model-Frozen Split Federated Learning for Resource-Efficient LLM Fine-Tuning over Heterogeneous Edge Devices

Collaborative fine-tuning of large language models (LLMs) at the network edge enables privacy-preserving and personalized generative intelligence, but is constrained by the limited memory and compute resources of edge devices. Traditional approaches, such as split federated learning (SFL), partition models between devices and a server, requiring iterative exchanges of activations/gradients while the server manages multiple submodels. This leads to substantial communication overhead, high activation storage costs, and severe server-side memory pressure when scaling to massive heterogeneous devices. In this paper, we propose FrozenSFL, a resource-efficient SFL framework for collaborative fine-tuning of LLMs across heterogeneous edge devices. At its core, FrozenSFL strictly freezes device-side submodels to eliminate both deviceside backpropagation and server-to-device gradient transmission, thereby removing on-device activation storage requirements and downlink communication overhead. Building upon this design, we improve the framework in two aspects. First, we design a layerwise LoRA aggregation strategy to handle structural heterogeneity introduced by device-specific model partitioning and enable efficient server-side updates. Second, we introduce a server-side feature caching mechanism that reuses invariant features, allowing subsequent training epochs to proceed offline and reducing redundant computation and communication. We implement the FrozenSFL system as an asynchronous multi-threaded pipeline, which serves as our real-world testbed for evaluation. On the device side, forward computation and data transmission are decoupled via a producer-consumer mechanism. On the server side, a first-in-first-out scheduling strategy serializes fine-tuning tasks across devices. A hybrid memory-disk streaming mechanism further buffers incoming data and ensures stable training under network variability. Extensive evaluations with representative models, including BERT-Base (0.11B), ViT-Large (0.3B), and LLaMA-2-7B (7B), validate the effectiveness of FrozenSFL. In particular, for LLaMA-2-7B, FrozenSFL can reduce device/server VRAM footprint by up to 14%/82.7% with only 1% accuracy loss, while achieving up to 6.7× fine-tuning speedup.

Xiao-Pei Chen +50
Scholarai-hardware

TD-dAJC: A 100 MS/s 2 pJ/Pixel Time-Domain Weight and Integrating-MAC-Based Direct Analog to MJPEG Compression for Video Sensor Nodes

Always-on, high-quality video sensors are increasingly required in wearables, surveillance, and Internet-of-Things (IoT) devices, where stringent energy constraints make efficient in-sensor compression necessary. These systems must support high-definition (HD) and 4 K video, yet conventional digital video compression pipelines consume significant power because high-speed analog-to-digital converters (ADCs) must digitize every pixel, and digital signal processing incurs large switching and memory overhead. This makes compression a major energy bottleneck in edge cameras. This work enables analog-domain MJPEG compression using a time-domain processing technique that represents discrete cosine transform (DCT) and quantization coefficients as pulsewidth–encoded time weights and performs multiply–accumulate (MAC) and division operations through current integration. The architecture employs time weights generated by a 32-phase sub-sampling PLL-based pulse generator, enabling MAC computation directly on analog pixel values before digitization, reducing the ADC workload. Fabricated in 65-nm CMOS, the proposed time-domain direct-analog-to-MJPEG compression (TD-dAJC) circuit achieves 100 MS/s throughput, the fastest reported among MJPEG implementations, and supports real-time HD and 4 K compression. The architecture consumes ~2 pJ/pixel, achieving <inline-formula> <tex-math notation="LaTeX">$13\times $ </tex-math></inline-formula> lower energy than switched-capacitor (SC) analog compressors and <inline-formula> <tex-math notation="LaTeX">$25\times $ </tex-math></inline-formula> lower energy than digital MJPEG pipelines. Compared to state-of-the-art analog in-memory compute MJPEG engines, TD-dAJC achieves <inline-formula> <tex-math notation="LaTeX">$2.8\times $ </tex-math></inline-formula> better energy efficiency and <inline-formula> <tex-math notation="LaTeX">$25\times $ </tex-math></inline-formula> higher throughput.

Harshit Naman +30IEEE Journal of Solid-State Circuits
Scholarai-hardware

Medusa: A Scalable Quantum-Inspired k-SAT Solver With Make and Break Feedback

Boolean satisfiability (SAT or <inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>-SAT, <inline-formula> <tex-math notation="LaTeX">$k \ge 3$ </tex-math></inline-formula>) is a binary combinatorial optimization problem (COP) that is challenging for traditional classical computing to address in polynomial time. Quantum and quantum-inspired systems have been touted to tackle these and similar nondeterministic polynomial-time (NP) problems. This work describes Medusa, a quantum-inspired <inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>-SAT solver capable of addressing problems of up to 200 variables and 1016 clauses. Implemented using analog and mixed-signal circuits, the solver benefits heavily from massive parallelism, continuous-time operation, and excellent energy efficiency, and advances the field of quantum-inspired, analog computing with best-in-class performance and energy efficiency. This performance is achieved with the introduction of three key architectural advancements: dual Make and Break feedback, a distributed logic feedback network, and multi-macro operation. The dual feedback mechanism improves the mean time to solution (TTS) for 50-variable 3-SAT problems by <inline-formula> <tex-math notation="LaTeX">$30\times $ </tex-math></inline-formula> when compared with a single feedback mechanism. The distributed logic in the feedback network implements all-to-all spin connectivity with arbitrary order interactions, enabling <inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>-SAT operation and supporting the scalability of the solver, which is further enhanced by multi-macro operation. Medusa, implemented in a 28 nm process, achieves a <inline-formula> <tex-math notation="LaTeX">$4.92~\mu $ </tex-math></inline-formula>s TTS for the popular SATLIB uf50-218 benchmark with a mean energy to solution (ETS) of 19.1 nJ. For the uf200-860 benchmark the solver achieves a 30.9 ms mean TTS, outperforming quantum annealing for problems of the same size.

L. Wormald +70IEEE Journal of Solid-State Circuits
Scholarai-hardware

An Intelligent Vision SoC With Fully Analog Sensing-to-Inference Pipeline and Linearity-Recovery RRAM Compute-in-Memory

Edge vision systems that integrate image sensing and neural-network inference on a single chip are constrained by inter-layer analog-to-digital converters (ADCs) and on-chip weight storage density, which together limit system efficiency and network scale. This article presents a fully analog vision system-on-chip (SoC) with 2-Mbit multi-level-cell (MLC) resistive random access memory (RRAM) that addresses both limitations through a continuous voltage–time datapath spanning pulsewidth-modulation (PWM) image sensing and multi-layer convolutional neural network (CNN) inference, eliminating inter-layer ADCs and off-chip memory access. The proposed linearity-recovery analog memory (LR-AMEM) compensates for the RC nonlinearity of direct charge-domain RRAM-based compute-in-memory (CIM), and a self-referenced threshold-variation-cancellation (TVC) mechanism suppresses process, voltage, and temperature (PVT) sensitivity. A parameterized controller supports flexible kernel configurations and dataflow patterns without data reordering. The 55-nm prototype achieves 11.07 pJ/(pixel <inline-formula> <tex-math notation="LaTeX">$\cdot $ </tex-math></inline-formula> frame) sensing efficiency and 345.54 TOPS/W system efficiency, a <inline-formula> <tex-math notation="LaTeX">$1.8\times $ </tex-math></inline-formula> to <inline-formula> <tex-math notation="LaTeX">$941.5\times $ </tex-math></inline-formula> improvement over prior intelligent imagers, and demonstrates feature extraction (7.44% RMSE), image classification (87.4% on CIFAR-10), and five-class object detection (20.5% mAP@0.5) on a single chip.

Zheng-Ke Yang +120IEEE Journal of Solid-State Circuits
Scholarai-hardware

A 28-nm Sign-Bit-Embedded Bit-Parallel SRAM Compute-in-Memory Macro With Stage-Wise-Enabled Transition-Counting-Lines and Accumulators for Edge-AI Devices

Computing-in-memory (CIM) architectures are promising for edge-AI devices, as they mitigate the von Neumann bottleneck and reduce data movement. Bit-parallel CIM macros with direct memory mapping and low toggle rate are well-suited for computing systems. However, severe routing congestion, inefficient signed computation, and unnecessary toggling of intermediate signals in local multiply-accumulate (MAC) circuits remain unsolved. This article presents a 28-nm bit-parallel SRAM CIM macro for efficient INT8 and BF16 MAC operations, featuring three techniques: 1) a transition-counting-line (TCL) that performs sparsity-adaptive, lossless bit-column compression by recording full-swing transitions derived from partial products, thereby mitigating routing congestion while preserving full precision; 2) a bus-based TCL that reuses sign-bit multiply-multiplex units (MMUs) to realize embedded sign extension with only 3% area overhead and no additional cycles; and 3) a shared ten-phase generator that enables TCLs and accumulations in a stage-wise manner to suppress spurious toggling, improving energy efficiency by 42%. Fabricated in 28-nm CMOS, the 32-kb macro occupies 0.072 mm2 and supports 64 input channels. The macro achieves a 4.1-ns access time at 0.9 V, with peak energy efficiencies of 106.85 TOPS/W in INT8 mode and 77.68 TFLOPS/W in BF16 mode, and peak area efficiencies of 1.73 TOPS/mm2 and 1.24 TFLOPS/mm2, respectively.

Yun-Long Liu +40IEEE Journal of Solid-State Circuits
Scholarai-hardware

A Valuewise Hybrid-Domain Compute-In-Memory Macro for Energy-Efficient Edge-AI Devices

This article presents a valuewise hybrid (VWH) compute-in-memory (CIM) macro that improves hybrid efficiency by <inline-formula> <tex-math notation="LaTeX">$1.75\times $ </tex-math></inline-formula> while simultaneously reducing computation error by <inline-formula> <tex-math notation="LaTeX">$2.02\times $ </tex-math></inline-formula> compared to prior bitwise hybrid schemes. The VWH approach performs accurate computations digitally for the most significant values (MSVs) in a small proportion, while the vast majority of the values are processed by lossy analog CIM (ACIM) macros to achieve high energy efficiency. To minimize latency and enhance the energy efficiency of the ACIM macros, a heterogeneous memory fabric (HMF) is introduced, which combines the benefits of static random access memory (SRAM)-CIM and dynamic random access memory (DRAM)-CIM. In addition, an asynchronous sparsity manager (ASM) enables adaptive throughput and power consumption based on sparsity. Fabricated using a 28-nm CMOS process, the prototype VWH-CIM macro achieves energy efficiencies of 470.5 TOPS/W in four-bit mode, 109.8 TOPS/W in eight-bit mode, and 48.5 TOPS/W in 12-bit mode.

Yuan-Zhe Zhao +60IEEE Journal of Solid-State Circuits
Scholarai-hardware

Multi-Mode In/Near-Pixel Processing CMOS Vision Sensor With Time-Memory-Based Temporal Derivative Computing

This article presents a temporal-derivative (TD) computing multi-mode vision sensor. The proposed sensor features a novel frame difference (FD) processing scheme that eliminates the need for frame memory by employing a checkerboard patch-based pixel architecture and column-shared processing circuits. This configuration efficiently detects patch-level temporal changes and generates 1-bit event signals (EVTs). Furthermore, an in-pixel time memory (TM) circuit is introduced to enable area- and power-efficient mixed-signal TD computing. The <inline-formula> <tex-math notation="LaTeX">$324~\times ~252$ </tex-math></inline-formula> prototype sensor was implemented in a 0.18-<inline-formula> <tex-math notation="LaTeX">$\mu $ </tex-math></inline-formula>m CMOS process. Measurement results demonstrate that the sensor consumes <inline-formula> <tex-math notation="LaTeX">$4.79~\mu $ </tex-math></inline-formula>W at 12.5 frames/s, achieving a FoM of 42.2 pJ/pix<inline-formula> <tex-math notation="LaTeX">$\cdot $ </tex-math></inline-formula>frame in the patch mode operation for FD-based motion detection and TD generation.

Seong-Min Ko +50IEEE Journal of Solid-State Circuits
Scholarai-hardware

RRAM-Based Computing-in-Memory FPGA Architecture for Low-Power Applications

Circuit design often faces a fundamental trade-off: ASICs offer peak efficiency at the cost of high design complexity and inflexibility, whereas FPGAs provide fast adaptability but suffer from high power consumption. In this work, we propose an RRAM-based FPGA architecture with a Reconfigurable Computing-in-Memory Logic Unit (RCLU), thereby achieving both flexibility, low cost, and low power consumption. By proposing a 2T2R structure that prevents RESET/SET operations in RRAM during logic computation, the proposed RCLU achieves significant improvements in delay and energy efficiency. In experimental benchmarks against conventional SRAM-FPGA and state-of-the-art RRAM-FPGA design, our architecture achieves power-delay product (PDP) improvements of up to 76.89% and 57.09%, respectively. Furthermore, to validate the low-power advantage of the proposed architecture, we implement ASCON, a representative lightweight IoT encryption algorithm. Simulation results show that our design achieves significantly lower energy per bit than conventional FPGAs and approaches ASIC-level efficiency, demonstrating strong potential as an energy-efficient IoT hardware solution.

Xiao-Chang Zhang +80IEEE Transactions on Circuits and Systems Part 1: Regular Papers