Muya Chang | Searcharxiv

arXiv2026

A 65-nm Privacy-Preserving Neuromorphic Encoder With 7.13-nJ Efficiency, 2.38-Mb/mm^2 Item-Memory Density, and Federated Learning Support

The increasing demand for privacy-preserving personal data analytics in smart assistants, wearable health monitors, and context-aware systems calls for hardware that is both energy-efficient and secure. This work presents a 65-nm privacy-preserving neuromorphic encoder that leverages transistor-level process variation as physically unclonable entropy for hyperdimensional computing. The proposed 2T-2T entropy cell enables compact, device-specific, and write-free item memory, allowing privacy-preserving bio-signal encoding without storing random basis vectors in conventional memory. The fabricated prototype achieves 7.13 nJ per encoding, 2.38 Mb/mm^2 item-memory density, 76.44 nJ per prediction, and 357.32 nJ per training update. It also supports in-situ decision-making, continual learning, and federated learning for multi-user deployment and cold-start personalization. Evaluations across bio-signal datasets demonstrate 93.2% accuracy on EMG and 96.1% accuracy on UCI-HAR, while reducing hypervector dimensionality by 14.3x compared with binary hyperdimensional computing. These results demonstrate an energy-efficient and privacy-preserving neuromorphic hardware platform for secure edge biomedical intelligence.

cs.AR↗

arXiv2025

Pretraining Large Language Models with NVFP4

Large Language Models (LLMs) today are powerful problem solvers across many domains, and they continue to get stronger as they scale in model size, training set size, and training set quality, as shown by extensive research and experimentation across the industry. Training a frontier model today requires on the order of tens to hundreds of yottaflops, which is a massive investment of time, compute, and energy. Improving pretraining efficiency is therefore essential to enable the next generation of even more capable LLMs. While 8-bit floating point (FP8) training is now widely adopted, transitioning to even narrower precision, such as 4-bit floating point (FP4), could unlock additional improvements in computational speed and resource utilization. However, quantization at this level poses challenges to training stability, convergence, and implementation, notably for large-scale models trained on long token horizons. In this study, we introduce a novel approach for stable and accurate training of large language models (LLMs) using the NVFP4 format. Our method integrates Random Hadamard transforms (RHT) to bound block-level outliers, employs a two-dimensional quantization scheme for consistent representations across both the forward and backward passes, utilizes stochastic rounding for unbiased gradient estimation, and incorporates selective high-precision layers. We validate our approach by training a 12-billion-parameter model on 10 trillion tokens -- the longest publicly documented training run in 4-bit precision to date. Our results show that the model trained with our NVFP4-based pretraining technique achieves training loss and downstream task accuracies comparable to an FP8 baseline. These findings highlight that NVFP4, when combined with our training approach, represents a major step forward in narrow-precision LLM training algorithms.

cs.CL↗