paperbot · PL 论文追踪

全部论文 待读 收藏

P2 PIM Is All You Need: A CXL-Enabled GPU-Free System for Large Language Model Inference

PIM 就是全部所需：一个基于 CXL 的免 GPU 大语言模型推理系统
ASPLOS Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 22025匹配度 1/10 · 该论文属于计算机体系结构/系统方向，研究用 CXL 内存扩展与近存储体计算构建免 GPU 的 LLM 推理硬件系统，关注吞吐、能效与 TCO，与用户关注的静态分析、形式化验证、浮点程序分析、贝叶斯程序分析及 LLM 与这些方向的交叉均无实质关联。虽然标签含“机器学习与程序语言”，但内容不涉及程序分析或 PL 方法。引用 74 inbox
Yufeng Gu, Alireza Khadem, Sumanth Umesh, Ning Liang, Xavier Servot, Onur Mutlu, Ravi Iyer, Reetuparna Das
语言与体系结构机器学习与程序语言
crossref: 2026-09-12dblp: 2026-09-12arxiv: 2026-09-12

AI 速览

一句话：提出 CENT——利用 CXL 内存扩展与近存储体计算构建免 GPU 的 LLM 推理系统，吞吐与能效显著优于 GPU 基线。

论文指出 LLM 自回归推理的运算强度远低于此前的编码器类模型与 CNN，且百万级 token 上下文带来的大键值缓存占用大量内存、限制批大小，因此推理本质上是内存带宽受限任务；而 GPU/TPU 等主流硬件却围绕计算吞吐优化，造成资源错配与高昂成本。为此作者提出 CENT，一个基于 CXL 的免 GPU LLM 推理系统：借助 CXL 内存扩展承载大规模模型参数，利用近存储体处理单元（near-bank PIM）提供高内存带宽，并通过可扩展的 CXL 网络支持设备间的点对点与集合通信原语，从而在多个 CXL 设备上实现多种并行策略来分布式部署 LLM。评估表明，在与 GPU 基线相同平均功耗、且采用其最大支持批大小的条件下，CENT 吞吐提升 2.3 倍、能耗降低 2.3 倍，每美元可生成的 token 数是 GPU 的 5.2 倍，显著降低总拥有成本。

问题
LLM 推理运算强度低、键值缓存庞大导致批大小受限，属内存带宽受限负载，而现有 GPU/TPU 面向计算吞吐优化，导致部署效率低、成本高。
方法
设计 CENT：用 CXL 内存扩展容纳大模型，用近存储体处理单元提供高带宽，并用可扩展 CXL 网络支持跨设备点对点与集合通信，以支撑多种模型并行策略。
评估
与采用最大支持批大小、平均功耗相近的 GPU 基线对比吞吐、能耗与总拥有成本（每美元 token 数）。
原文摘要（Abstract）

Large Language Model (LLM) inference uses an autoregressive manner to generate one token at a time, which exhibits notably lower operational intensity compared to earlier Machine Learning (ML) models such as encoder-only transformers and Convolutional Neural Networks. At the same time, LLMs possess large parameter sizes and use key-value caches to store context information. Modern LLMs support context windows with up to 1 million tokens to generate versatile text, audio, and video content. A large key-value cache unique to each prompt requires a large memory capacity, limiting the inference batch size. Both low operational intensity and limited batch size necessitate a high memory bandwidth. However, contemporary hardware systems for ML model deployment, such as GPUs and TPUs, are primarily optimized for compute throughput. This mismatch challenges the efficient deployment of advanced LLMs and makes users to pay for expensive compute resources that are poorly utilized for the memory-bound LLM inference tasks. We propose CENT, a CXL-ENabled GPU-Free sysTem for LLM inference, which harnesses CXL memory expansion capabilities to accommodate substantial LLM sizes, and utilizes near-bank processing units to deliver high memory bandwidth, eliminating the need for expensive GPUs. CENT exploits a scalable CXL network to support peer-to-peer and collective communication primitives across CXL devices. We implement various parallelism strategies to distribute LLMs across these devices. Compared to GPU baselines with maximum supported batch sizes and similar average power, CENT achieves 2.3x higher throughput and consumes 2.3x less energy. CENT reduces the Total Cost of Ownership (TCO), generating 5.2x more tokens per dollar than GPUs.

链接与引用

DOI 原文 · arXiv · PDF（开放获取） · DBLP

BibTeX 
@inproceedings{abs-2502-07578,
  title = {PIM Is All You Need: A CXL-Enabled GPU-Free System for Large Language Model Inference},
  author = {Yufeng Gu and Alireza Khadem and Sumanth Umesh and Ning Liang and Xavier Servot and Onur Mutlu and Ravi Iyer and Reetuparna Das},
  booktitle = {ASPLOS},
  year = {2025},
  doi = {10.1145/3676641.3716267}
}