LUT-SSM: A 99.3TFLOPS/W LUT-Based State-Space Model Accelerator Using Energy-Efficient Element-Wise Layer Fusion and LUT-Friendly Weight-Only Quantization

📋 论文概要

提出一种基于查找表（LUT）的状态空间模型（SSM）加速器LUT-SSM，通过多对多LUT-ACC和逐元素层融合技术高效支持INT-FP的通用矩阵乘法和顺序逐元素运算，解决SSM部署中的存储访问和能效问题。在28nm FD-SOI工艺上实现99.3TFLOPS/W峰值能效。

💡 主要创新点

• 提出基于LUT的many-to-many累加器（LUT-ACC），高效支持SSM中的INT-FP矩阵乘法。
• 引入逐元素（EW）层融合技术，减少外部存储访问并提升能效。
• 整体架构支持SSM块中的顺序逐元素运算，实现高能效推理。

🏷 关键词

📄 原文摘要

Ulsan National Institute of Science and Technology, Ulsan, Korea, 4Naver Cloud, Seongnam, Korea *Equally Credited Authors (ECAs) 1 3 Abstract State-space models (SSMs) and weight-only quantization alleviate huge external memory access with a minimal accuracy degradation. To support both efficiently, we propose LUTSSM, a LUT-based SSM accelerator with a many-to-many LUT-ACC and an element-wise (EW) layer fusion. LUT-SSM efficiently supports INT-FP GEMM and sequential EW operations in the SSM blocks. Fabricated in 28nm FD-SOI, LUT-SSM achieves 99.3TFLOPS/W peak efficiency and 96.1TFLOPS/W on Mamba 1.4B, improving energy efficiency by 1.28× over recent designs. Recently, attention-based models [1] have been deployed in a variety of applications, e.g., AI agents, autonomous driving, robotics, and healthcare. However, they suffer from significant external memory access (EMA) for weights and intermediate caches, leading to

👥 作者与机构

Sunwoo Yoo*1,2, Dongyun Kam*3, Gunho Park4, Soonhyun Kwon1, Dongsoo Lee4, Youngjoo Lee1

Korea Advanced Institute of Science and Technology, Daejeon, Korea, 2Pohang University of Science and Technology, Pohang, Korea,