Saved in:
Bibliographic Details
Main Authors: Price, Daniel, Vellaisamy, Prabhu, Shen, John, Wu, Di
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2601.10823
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911421632086016
author Price, Daniel
Vellaisamy, Prabhu
Shen, John
Wu, Di
author_facet Price, Daniel
Vellaisamy, Prabhu
Shen, John
Wu, Di
contents Value level parallelism (VLP) has been proposed to improve the efficiency of large-batch, low-precision general matrix multiply (GEMM) between symmetric activations and weights. In transformer based large language models (LLMs), there exist more sophisticated operations beyond activation-weight GEMM. In this paper, we explore how VLP benefits LLMs. First, we generalize VLP for nonlinear approximations, outperforming existing nonlinear approximations in end-to-end LLM accuracy, performance, and efficiency. Our VLP approximation follows a value-centric approach, where important values are assigned with greater accuracy. Second, we optimize VLP for small-batch GEMMs with asymmetric inputs efficiently, which leverages timely LLM optimizations, including weight-only quantization, key-value (KV) cache quantization, and group query attention. Finally, we design a new VLP architecture, Mugi, to encapsulate the innovations above and support full LLM workloads, while providing better performance, efficiency and sustainability. Our experimental results show that Mugi can offer significant improvements on throughput and energy efficiency, up to $45\times$ and $668\times$ for nonlinear softmax operations, and $2.07\times$ and $3.11\times$ for LLMs, and also decrease operational carbon for LLM operation by $1.45\times$ and embodied carbon by $1.48\times$.
format Preprint
id arxiv_https___arxiv_org_abs_2601_10823
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Mugi: Value Level Parallelism For Efficient LLMs
Price, Daniel
Vellaisamy, Prabhu
Shen, John
Wu, Di
Machine Learning
Hardware Architecture
Value level parallelism (VLP) has been proposed to improve the efficiency of large-batch, low-precision general matrix multiply (GEMM) between symmetric activations and weights. In transformer based large language models (LLMs), there exist more sophisticated operations beyond activation-weight GEMM. In this paper, we explore how VLP benefits LLMs. First, we generalize VLP for nonlinear approximations, outperforming existing nonlinear approximations in end-to-end LLM accuracy, performance, and efficiency. Our VLP approximation follows a value-centric approach, where important values are assigned with greater accuracy. Second, we optimize VLP for small-batch GEMMs with asymmetric inputs efficiently, which leverages timely LLM optimizations, including weight-only quantization, key-value (KV) cache quantization, and group query attention. Finally, we design a new VLP architecture, Mugi, to encapsulate the innovations above and support full LLM workloads, while providing better performance, efficiency and sustainability. Our experimental results show that Mugi can offer significant improvements on throughput and energy efficiency, up to $45\times$ and $668\times$ for nonlinear softmax operations, and $2.07\times$ and $3.11\times$ for LLMs, and also decrease operational carbon for LLM operation by $1.45\times$ and embodied carbon by $1.48\times$.
title Mugi: Value Level Parallelism For Efficient LLMs
topic Machine Learning
Hardware Architecture
url https://arxiv.org/abs/2601.10823