Inspiration
Basically my work has been in performance engineering mostly and I have been fascinated by how that works in gpu world I have profiled in cpu using different profilers but new to gpu performance
What it does
gpu-perf basically does perf for gpu the
How we built it
Codex heavily involved (i am hoping when i am able to test it with nvidia I won't have much to do) using C++ and ebpf (this needs to be refined)
Challenges we ran into
Getting actionable insights from the kernel I would try to set it up in a way that doesn't bother the user privilege but that depends on linux more and what we can get without PERFMON
Accomplishments that we're proud of
I'd say I got to understand architecture of my machine better, i utilized sycl to test and run my intel integrated gpu which has opened up doors to new knowledge
What we learned
It is difficult to do it cross vendor and resource constrained. I thought I will be able to run tests against cloud gpus but they are also difficult to do properly
What's next for gpu_perf
Some testing on nvidia and amd, reducing footprint on ram and disk more actionable insights.
Improve binary size and memory consumption utilize
Built With
- c++
- ebpf
Log in or sign up for Devpost to join the conversation.