Clemens Schaefer

Clemens Schaefer, Gil Tabak

Microscaling is a critical technique for preserving the quality of Large Language Models (LLMs) quantized to ultra-low precision formats. Intuitively, finer block sizes should yield lower quantization error; however, a paradox recently identified in the literature demonstrates that standard abs-max scaling can actually degrade model quality as block sizes shrink. In this work, we investigate the underlying mechanics of this phenomenon. We demonstrate that this degradation is not an inherent limitation of finer granularity, but is primarily driven by heavy-tailed tensor distributions interacting poorly with the coarse upper quantization bins of the FP4 element format. Specifically, we show that i) preventing the scaling factor from underflowing to zero mitigates localized errors, ii) targeted algorithmic interventions like the 4-over-6 methodology effectively correct the quantization geometry for large elements, and iii) a brute-force search establishes an optimal baseline, confirming that the theoretical Mean Squared Error (MSE) strictly improves with finer block sizes. Ultimately, our findings reveal a valuable interchangeability: applying the correct algorithmic recipe allows standard, hardware-compliant formats (like OCP E4M3) to match the performance of custom, wider-exponent formats (like UE5M3). We validate these results across several large language models, fully resolving the block size paradox and achieving robust downstream perplexity improvements.

Ibrahim Ahmed, Clemens Schaefer, Gil Tabak et al.

While Large Language Models (LLMs) have become highly influential, their enormous scale presents significant deployment challenges. Efficiently serving these models typically requires distributing them across numerous accelerator devices, which introduces substantial performance overhead from inter-device communication (collectives). While model quantization has been widely adopted to reduce the memory and compute requirements of LLM weights and activations with minimal quality impact, applying quantization directly to collectives like AllReduce is inherently difficult due to the inter-device summation involved, which can lead to numerical instability or significant error accumulation. In this work, we present a native dynamic block-wise efficient quantized AllReduce within the XLA compiler for TPUs (EQuARX). By using TPU-friendly quantization and deep pipelining of communication and compute, EQuARX with int8 precision achieves a 1.8X speedup over baseline BF16 AllReduce across various network topologies. Furthermore, EQuARX accelerates the prefill stage of Gemma 3 27B by 1.25X and Gemma 3 12B by 1.1X, respectively, with small to negligible impact on quality.

Weier Wan, Rajkumar Kubendran, Clemens Schaefer et al.

Realizing today's cloud-level artificial intelligence functionalities directly on devices distributed at the edge of the internet calls for edge hardware capable of processing multiple modalities of sensory data (e.g. video, audio) at unprecedented energy-efficiency. AI hardware architectures today cannot meet the demand due to a fundamental "memory wall": data movement between separate compute and memory units consumes large energy and incurs long latency. Resistive random-access memory (RRAM) based compute-in-memory (CIM) architectures promise to bring orders of magnitude energy-efficiency improvement by performing computation directly within memory. However, conventional approaches to CIM hardware design limit its functional flexibility necessary for processing diverse AI workloads, and must overcome hardware imperfections that degrade inference accuracy. Such trade-offs between efficiency, versatility and accuracy cannot be addressed by isolated improvements on any single level of the design. By co-optimizing across all hierarchies of the design from algorithms and architecture to circuits and devices, we present NeuRRAM - the first multimodal edge AI chip using RRAM CIM to simultaneously deliver a high degree of versatility for diverse model architectures, record energy-efficiency $5\times$ - $8\times$ better than prior art across various computational bit-precisions, and inference accuracy comparable to software models with 4-bit weights on all measured standard AI benchmarks including accuracy of 99.0% on MNIST and 85.7% on CIFAR-10 image classification, 84.7% accuracy on Google speech command recognition, and a 70% reduction in image reconstruction error on a Bayesian image recovery task. This work paves a way towards building highly efficient and reconfigurable edge AI hardware platforms for the more demanding and heterogeneous AI applications of the future.