Skip to content

In the presence of multiple OpenCL types, Milkyway runs on the wrong GPU #132

@piedar

Description

@piedar

A machine has multiple OpenCL GPUs, which BOINC reports like this.

OpenCL: NVIDIA GPU 0: NVIDIA GeForce GTX 1050 (driver version 515.65.01, device version OpenCL 3.0 CUDA, 2000MB, 2000MB available, 1911 GFLOPS peak)
OpenCL: AMD/ATI GPU 0: AMD Radeon Vega 8 Graphics (raven, LLVM 14.0.6, DRM 3.42, 5.15.6 (driver version 22.1.7, device version OpenCL 1.1 Mesa 22.1.7, 3072MB, 3072MB available, 704 GFLOPS peak)

BOINC downloads tasks for both NVIDIA and AMD, but Milkyway runs both tasks on the NVIDIA GPU (as shown by nvidia-smi) and leaves the AMD device idle. According to BOINC/boinc#1641 Milkyway should be using gpu_opencl_dev_index rather than the raw numeric device index.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions