arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2601.13628cs.AR
收藏

PRIMAL:基于内存计算的低秩适应LLM推理加速器

PRIMAL: Processing-In-Memory Based Low-Rank Adaptation for LLM Inference Accelerator

National University of Singapore(新加坡国立大学)

Yue Jiet Chong, Yimin Wang, Zhen Wu, Xuanyao Fong

AI总结:

PRIMAL通过内存计算和低秩适应技术,实现LLM推理加速,相比NVIDIA H100在吞吐量和能效上有显著提升。

摘要:

本文提出PRIMAL,一种基于内存计算(PIM)的大型语言模型(LLM)推理加速器,结合低秩适应(LoRA)。PRIMAL整合了异构的内存计算处理元件(PEs),通过2D网格互连的计算网络(IPCN)连接。一种新颖的SRAM重编程和电源门控(SRPG)方案通过重配置与计算和门控空闲资源的重叠,实现了流水线化的LoRA更新和亚线性功率缩放。PRIMAL采用优化的空间映射和数据流协调,以最小化通信开销,并在LoRA秩8(Q,V)上实现Llama-13B的1.5倍吞吐量和25倍能效,优于NVIDIA H100。

英文摘要:

This paper presents PRIMAL, a processing-in-memory (PIM) based large language model (LLM) inference accelerator with low-rank adaptation (LoRA). PRIMAL integrates heterogeneous PIM processing elements (PEs), interconnected by 2D-mesh inter-PE computational network (IPCN). A novel SRAM reprogramming and power gating (SRPG) scheme enables pipelined LoRA updates and sub-linear power scaling by overlapping reconfiguration with computation and gating idle resources. PRIMAL employs optimized spatial mapping and dataflow orchestration to minimize communication overhead, and achieves $1.5\times$ throughput and $25\times$ energy efficiency over NVIDIA H100 with LoRA rank 8 (Q,V) on Llama-13B.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑