Announcement

Collapse
No announcement yet.

Radeon ROCm 2.0 OpenCL Benchmarks With Linux 5.0 On Ubuntu 18.10 vs. NVIDIA's Linux Driver

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • max0x7ba
    replied
    In terms of Machine Learning performance the only framework worth benchmarking and whose results matter is Tensorflow.

    GitHub stats: PlaidML - 1.9k stars, Tensorflow - 121k stars.

    Very few people heard or care about PlaidML.

    Leave a comment:


  • coder
    replied
    Originally posted by smitty3268 View Post
    Pretty much, yeah. I think they're bumping some clock speeds a small amount and limiting the fp64 rate vs the pro card.
    It's 1/4th of the MI60, whereas Vega 64 is about half that. So, it'll net you more than 1 TFLOPS of fp64.

    I guess that's something.

    Leave a comment:


  • coder
    replied
    Originally posted by s_j_newbury View Post
    No double precision benchmarks?
    Michael has the wrong hardware, if you care about fp64.

    Nvidia's Titan V is rated at 6.14 TFLOPS, while AMD's MI60 is rated at 7.36 TFLOPS. Anything consumer-grade is a toy, by comparison.

    Leave a comment:


  • coder
    replied
    Originally posted by AndyChow View Post
    Radeon VII is MI50 made for consumers, right? I would like to see how it performs these tests, since it`s supposed to be a platform for ML/AI.
    Well, it > doubles memory bandwidth, but has only 60 CUs (instead of 64) and just runs them at higher clocks. So, unless Michael enables 8-bit or 4-bit options for inferencing, probably don't expect too much. Certainly, no more than about 2x (in memory-bound workloads).

    And it's still no match for Nvidia's Tensor cores (which these benchmarks don't seem to be using).
    Last edited by coder; 03 February 2019, 02:01 PM.

    Leave a comment:


  • pal666
    replied
    so what is wrong with first batch of tests where amd cards perform much worse than on rest of tests?

    Leave a comment:


  • cb88
    replied
    kernel latency is also way higher on amdocl...

    Basically there is mo reason to not use RoCm on amd cards... it also remains to be seen If the rate limited fp64 can be worked around in the open drivers.

    Leave a comment:


  • Happy Heyoka
    replied
    Originally posted by ObiWan View Post
    For those curious what a Vega 56 (165 Watt powerlimit) does with the amdgpu pro opencl backend, instead of ROCm

    https://openbenchmarking.org/result/...SP-1901305SP07
    So if I'm reading that right, comparable to the ROCm on almost everything but memory performance?
    I wonder what's going on there and how much whatever that is is contributing to the slight discrepancies here and there...?

    Leave a comment:


  • smitty3268
    replied
    Originally posted by AndyChow View Post
    Radeon VII is MI50 made for consumers, right?
    Pretty much, yeah. I think they're bumping some clock speeds a small amount and limiting the fp64 rate vs the pro card.

    Leave a comment:


  • s_j_newbury
    replied
    No double precision benchmarks?

    Leave a comment:


  • ObiWan
    replied
    For those curious what a Vega 56 (165 Watt powerlimit) does with the amdgpu pro opencl backend, instead of ROCm

    https://openbenchmarking.org/result/...SP-1901305SP07

    Leave a comment:

Working...
X