Optimizing Attention on GPUs by Exploiting GPU Architectural NUMA Effects

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Choudhary, Mansi, Sangaiah, Karthik, Singh, Sonali, Osama, Muhammad, Wills, Lisa Wu, Dasika, Ganesh
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912686726447104
author Choudhary, Mansi
Sangaiah, Karthik
Singh, Sonali
Osama, Muhammad
Wills, Lisa Wu
Dasika, Ganesh
author_facet Choudhary, Mansi
Sangaiah, Karthik
Singh, Sonali
Osama, Muhammad
Wills, Lisa Wu
Dasika, Ganesh
contents The rise of disaggregated AI GPUs has exposed a critical bottleneck in large-scale attention workloads: non-uniform memory access (NUMA). As multi-chiplet designs become the norm for scaling compute capabilities, memory latency and bandwidth vary sharply across compute regions, undermining the performance of traditional GPU kernel scheduling strategies that assume uniform memory access. We identify how these NUMA effects distort locality in multi-head attention (MHA) and present Swizzled Head-first Mapping, a spatially-aware scheduling strategy that aligns attention heads with GPU NUMA domains to exploit intra-chiplet cache reuse. On AMD's MI300X architecture, our method achieves up to 50% higher performance over state-of-the-art attention algorithms using conventional scheduling techniques and sustains consistently high L2 cache hit rates of 80-97%. These results demonstrate that NUMA-aware scheduling is now fundamental to achieving full efficiency on next-generation disaggregated GPUs, offering a path forward for scalable AI training and inference.
format Preprint
id arxiv_https___arxiv_org_abs_2511_02132
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Optimizing Attention on GPUs by Exploiting GPU Architectural NUMA Effects
Choudhary, Mansi
Sangaiah, Karthik
Singh, Sonali
Osama, Muhammad
Wills, Lisa Wu
Dasika, Ganesh
Hardware Architecture
Distributed, Parallel, and Cluster Computing
Machine Learning
Performance
The rise of disaggregated AI GPUs has exposed a critical bottleneck in large-scale attention workloads: non-uniform memory access (NUMA). As multi-chiplet designs become the norm for scaling compute capabilities, memory latency and bandwidth vary sharply across compute regions, undermining the performance of traditional GPU kernel scheduling strategies that assume uniform memory access. We identify how these NUMA effects distort locality in multi-head attention (MHA) and present Swizzled Head-first Mapping, a spatially-aware scheduling strategy that aligns attention heads with GPU NUMA domains to exploit intra-chiplet cache reuse. On AMD's MI300X architecture, our method achieves up to 50% higher performance over state-of-the-art attention algorithms using conventional scheduling techniques and sustains consistently high L2 cache hit rates of 80-97%. These results demonstrate that NUMA-aware scheduling is now fundamental to achieving full efficiency on next-generation disaggregated GPUs, offering a path forward for scalable AI training and inference.
title Optimizing Attention on GPUs by Exploiting GPU Architectural NUMA Effects
topic Hardware Architecture
Distributed, Parallel, and Cluster Computing
Machine Learning
Performance
url https://arxiv.org/abs/2511.02132