GPUWorks open-sources a 40MB inference platform for mixed-chip clusters
AIGPUWorks, a lightweight enterprise AI inference platform, was open-sourced under the Apache-2.0 license on November 4 and runs NVIDIA, AMD, Huawei Ascend, T-Head PPU and Hygon DCU accelerators in one pool. According to the team's own tests, a single server managed 1,003 GPU servers and restored a sleeping model in about four seconds, versus 16 to 17.5 minutes for a cold start from NFS storage.
AIWhy it matters.
The source gives concrete benchmarks for multi-vendor scheduling and model wake-up, showing how mixed accelerator clusters can be managed and how cold-start time was cut.
Source details
Pandaily · contact@pandaily.com (Pandaily) · pandaily.comPublished · added here
Original title: GPUWorks Open-Sources 40MB Inference Platform Running Ascend, PPU, Hygon and NVIDIA in One Cluster