Page MenuHomePhabricator

GPU resources for Labs
Open, Needs TriagePublic

Description

Reading, research, and the scoring team want to have GPU resources in Labs. What's that going to take?

Event Timeline

Halfak created this task.Feb 27 2017, 5:37 PM
Restricted Application added a project: Cloud-Services. · View Herald TranscriptFeb 27 2017, 5:37 PM
Restricted Application added a subscriber: Aklapper. · View Herald Transcript
hashar added a subscriber: hashar.Mar 1 2017, 6:00 PM

Linux Kernel 4.10 added virtual GPU granted you have Intel Haswell with Intel Graphics https://kernelnewbies.org/Linux_4.10#head-6b7d1bd2f7247da9c87798a91e10302b0cc32331 . Would potentially let a KVM guest to access the host GPU.

dr0ptp4kt added a comment.EditedMar 1 2017, 6:01 PM

I wanted to note something uncovered while looking into this today.

TensorFlow CUDA support on Nvidia would currently require encumbered libraries, including in shared library space, so CUDA specific tie-ins (would be beneficial for TensorFlow) aren't a viable option on our infrastructure.

Now, that isn't to say there isn't OpenCL support for Nvidia - there is, according to https://developer.nvidia.com/opencl and http://www.nvidia.com/download/driverResults.aspx/115283/en-us (Tesla P100 would be a quality GPU). Presuming that driver would be acceptable, there would be an opportunity to ride on OpenCL technology (looks like it hinges on C/C++, or similar bindings).

bd808 moved this task from Triage to Stalled on the Cloud-Services board.Mar 26 2017, 10:45 PM
hashar removed a subscriber: hashar.Jun 16 2017, 11:27 AM