Tweet
Originally posted by pszilard
View Post
I really wish there were higher quality and more representative compute/OpenCL benchmarks. That's a pretty solid chunk of the Linux audience, I'd say and the current set of benchmarks are little and ill conceived IMHO.
- This GpuTest is pretty outdated OpenGL compute shader stuff, I doubt it's very representative for compute.
- Mixbench is, well, as its author describes: "The purpose of this benchmark tool is to evaluate performance bounds of GPUs on mixed operational intensity kernels. The executed kernel is customized on a range of different operational intensity values."
Without seeing a _curve_ of flops/byte vs performance, not even the flops/byte at which the tests are run, those numbers are meaningless. Even a simple peak flop rate (with the arithmetic intensity at which this is reached) would be more meaningful than just a single number.
- Last, the SHOC MD5 is frankly a niche benchmark that's heavy on integer and bitwise ops so it's special and that load is frankly not too interesting for most compute use-cases.
I'm baffled why aren't any of the standard/classic computational/HPC benchmarks included that are the de-facto tools to characterize the performance of a processor/compute architecture. These are in fact a the foundation of SHOC, included in the "Level 0" (device memory bw, max flops, etc.) and "Level 1" tests (FFT, scan, reduction, gemm, sort, triad, etc.): https://github.com/vetter/shoc/wiki
Michael, can you please take this feedback and consider it seriously. These GPUs are not anymore gamers-only anymore an I'm sure there are plenty of people in the technical computing/HPC community interested (I know at least a dozen
) in more thorough and representative compute benchmarks.
- This GpuTest is pretty outdated OpenGL compute shader stuff, I doubt it's very representative for compute.
- Mixbench is, well, as its author describes: "The purpose of this benchmark tool is to evaluate performance bounds of GPUs on mixed operational intensity kernels. The executed kernel is customized on a range of different operational intensity values."
Without seeing a _curve_ of flops/byte vs performance, not even the flops/byte at which the tests are run, those numbers are meaningless. Even a simple peak flop rate (with the arithmetic intensity at which this is reached) would be more meaningful than just a single number.
- Last, the SHOC MD5 is frankly a niche benchmark that's heavy on integer and bitwise ops so it's special and that load is frankly not too interesting for most compute use-cases.
I'm baffled why aren't any of the standard/classic computational/HPC benchmarks included that are the de-facto tools to characterize the performance of a processor/compute architecture. These are in fact a the foundation of SHOC, included in the "Level 0" (device memory bw, max flops, etc.) and "Level 1" tests (FFT, scan, reduction, gemm, sort, triad, etc.): https://github.com/vetter/shoc/wiki
Michael, can you please take this feedback and consider it seriously. These GPUs are not anymore gamers-only anymore an I'm sure there are plenty of people in the technical computing/HPC community interested (I know at least a dozen
) in more thorough and representative compute benchmarks.
Comment