Scalable Processing-Near-Memory for 1M-Token LLM Inference: CXL-Enabled KV-Cache Management Beyond GPU Limits

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Kim, Dowon, Lee, MinJae, Kim, Janghyeon, Kwon, HyuckSung, Jeong, Hyeonggyu, Park, Sang-Soo, Yoon, Minyong, Roh, Si-Dong, Kwon, Yongsuk, So, Jinin, Choi, Jungwook
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909880940494848
author Kim, Dowon
Lee, MinJae
Kim, Janghyeon
Kwon, HyuckSung
Jeong, Hyeonggyu
Park, Sang-Soo
Yoon, Minyong
Roh, Si-Dong
Kwon, Yongsuk
So, Jinin
Choi, Jungwook
author_facet Kim, Dowon
Lee, MinJae
Kim, Janghyeon
Kwon, HyuckSung
Jeong, Hyeonggyu
Park, Sang-Soo
Yoon, Minyong
Roh, Si-Dong
Kwon, Yongsuk
So, Jinin
Choi, Jungwook
contents The expansion of context windows in large language models (LLMs) to multi-million tokens introduces severe memory and compute bottlenecks, particularly in managing the growing Key-Value (KV) cache. While Compute Express Link (CXL) enables non-eviction frameworks that offload the full KV-cache to scalable external memory, these frameworks still suffer from costly data transfers when recalling non-resident KV tokens to limited GPU memory as context lengths increase. This work proposes scalable Processing-Near-Memory (PNM) for 1M-Token LLM Inference, a CXL-enabled KV-cache management system that coordinates memory and computation beyond GPU limits. Our design offloads token page selection to a PNM accelerator within CXL memory, eliminating costly recalls and enabling larger GPU batch sizes. We further introduce a hybrid parallelization strategy and a steady-token selection mechanism to enhance compute efficiency and scalability. Implemented atop a state-of-the-art CXL-PNM system, our solution delivers consistent performance gains for LLMs with up to 405B parameters and 1M-token contexts. Our PNM-only offloading scheme (PNM-KV) and GPU-PNM hybrid with steady-token execution (PnG-KV) achieve up to 21.9x throughput improvement, up to 60x lower energy per token, and up to 7.3x better total cost efficiency than the baseline, demonstrating that CXL-enabled multi-PNM architectures can serve as a scalable backbone for future long-context LLM inference.
format Preprint
id arxiv_https___arxiv_org_abs_2511_00321
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Scalable Processing-Near-Memory for 1M-Token LLM Inference: CXL-Enabled KV-Cache Management Beyond GPU Limits
Kim, Dowon
Lee, MinJae
Kim, Janghyeon
Kwon, HyuckSung
Jeong, Hyeonggyu
Park, Sang-Soo
Yoon, Minyong
Roh, Si-Dong
Kwon, Yongsuk
So, Jinin
Choi, Jungwook
Hardware Architecture
Artificial Intelligence
The expansion of context windows in large language models (LLMs) to multi-million tokens introduces severe memory and compute bottlenecks, particularly in managing the growing Key-Value (KV) cache. While Compute Express Link (CXL) enables non-eviction frameworks that offload the full KV-cache to scalable external memory, these frameworks still suffer from costly data transfers when recalling non-resident KV tokens to limited GPU memory as context lengths increase. This work proposes scalable Processing-Near-Memory (PNM) for 1M-Token LLM Inference, a CXL-enabled KV-cache management system that coordinates memory and computation beyond GPU limits. Our design offloads token page selection to a PNM accelerator within CXL memory, eliminating costly recalls and enabling larger GPU batch sizes. We further introduce a hybrid parallelization strategy and a steady-token selection mechanism to enhance compute efficiency and scalability. Implemented atop a state-of-the-art CXL-PNM system, our solution delivers consistent performance gains for LLMs with up to 405B parameters and 1M-token contexts. Our PNM-only offloading scheme (PNM-KV) and GPU-PNM hybrid with steady-token execution (PnG-KV) achieve up to 21.9x throughput improvement, up to 60x lower energy per token, and up to 7.3x better total cost efficiency than the baseline, demonstrating that CXL-enabled multi-PNM architectures can serve as a scalable backbone for future long-context LLM inference.
title Scalable Processing-Near-Memory for 1M-Token LLM Inference: CXL-Enabled KV-Cache Management Beyond GPU Limits
topic Hardware Architecture
Artificial Intelligence
url https://arxiv.org/abs/2511.00321