# DxPU: Large Scale Disaggregated GPU Pools in the Datacenter

**Authors:** Bowen He, Xiao Zheng, Yuan Chen, Weinan Li, Yajin Zhou, Xin Long, Pengcheng Zhang, Xiaowei Lu, Linquan Jiang, Qiang Liu, Dennis Cai, Xiantao Zhang  
**arXiv:** [2310.04648](https://arxiv.org/abs/2310.04648)  
**Date:** 2023-10-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.04648v1/

## Key points

- Increasing demand and convenience have led to rapid adoption of GPUs
- Traditional methods of attaching GPUs to host servers are inefficient
- GPU disaggregation technique proposed to address this issue
- DxPU is a new implementation of datacenter-scale GPU disaggregation
- DxPU efficiently solves previous problems and can allocate as many GPU nodes as needed
- Performance model developed for AI workloads and experiments conducted to evaluate system performance overhead
- Results show that DxPU incurs less than 10% overhead compared to native GPU servers in most scenarios
- DxPU introduced as an efficient solution for large-scale GPU disaggregation in datacenters, with promising results in terms of performance overhead.

## Summary

The increasing demand for  and the convenience of  have led to a rapid adoption of . However, traditional methods of attaching GPUs to host servers as PCIe devices are highly inefficient. To address this issue, a technique called  has been proposed, which involves aggregating GPUs into a pool and allocating them based on user demands. This paper presents a new implementation of datacenter-scale GPU disaggregation called DxPU, which efficiently solves previous problems and can flexibly allocate as many GPU nodes as needed. The authors also develop a performance model specifically for AI workloads and conduct experiments to evaluate the performance overhead caused by their system. Results show that DxPU incurs an overhead of less than 10% compared to native GPU servers in most scenarios. Overall, this paper introduces DxPU as an efficient solution for large-scale GPU disaggregation in datacenters, addressing previous limitations and providing promising results in terms of performance overhead.