GPUInspector.jl
GPUInspector.jl copied to clipboard
pc2
Inspecting GPUs with Julia
This is an initial attempt to support Nvidia Hopper GPUs, opening as draft because lots of thing still don't work. For example, theoretical peakflops for tensor cores is wrong, it...
* Requires https://github.com/JuliaGPU/AMDGPU.jl/pull/471 * Requires https://github.com/JuliaGPU/AMDGPU.jl/pull/472 TODOs (at least some of them 😉) * [x] AMD GPU CI * [ ] `memory_bandwidth_saxpy_*` (runs but values are suspicious/wrong) * [ ]...
* Binary? -> https://github.com/JuliaLang/PackageCompiler.jl * Script that installs and runs everything? * ...
Once https://github.com/pc2/GPUInspector.jl/issues/18 is done, add AMD GPU support in separate pkg extension.
Currently they don't look good. Screenshots would probably be better.
* [ ] Propper announcement on Julia discourse * [ ] Catchy "experience" blog post (with unicode plots etc.) * [ ] document usage for non-Julia users
https://app.codecov.io/gh/pc2/GPUInspector.jl
https://github.com/pc2/GPUInspector.jl/blob/89120d05ff3d86c6646ce1a52d08db04f0a8ed60/src/peakflops_gpu.jl#L56
Maybe we should also implement some cuda-aware MPI tests. These tests could .. - check weather the cuda-aware MPI installation actually works properly. - benchmark gpugpu communication between nodes.
Currently we query the device/compute capability to figure out * which data types are supported by tensor cores (https://github.com/pc2/GPUInspector.jl/blob/f5e17e589c2403222d616f7539dd69b6131bbfe4/src/peakflops_gpu.jl#L61) * the number of cuda cores (https://github.com/pc2/GPUInspector.jl/blob/f5e17e589c2403222d616f7539dd69b6131bbfe4/src/gpuinfo.jl#L248) * etc.? Can we...