1
$\begingroup$

I'm looking to build a PC for deep learning will tensoflow work on amd GPU with the same speed as on nvidia ones as amd doesn't have tensorcore or cuda cores but it will have 16gb of hbm vram his much do the tensor cores impact training

$\endgroup$
  • $\begingroup$ The problem with the Radeon VII which will released in Feb 2019 is the low amount of RAM. Only 16 gb hbm2 memory is builtin. That is not enough for serious deeplearning. The better idea is to buy the Tesla M10 model which provides 32 gb of RAM. More is better. $\endgroup$ – Manuel Rodriguez Jan 10 at 20:53
  • $\begingroup$ Tensorflow and Pytorch have now ROCm suppport for AMD GPUs, although I have not found any benchmarks. $\endgroup$ – TFreitas Feb 12 at 10:46
1
$\begingroup$

Tensorflow only works with Nvidia gpus. There are solutions that allegedly also work with AMD, but it is certainly easier to set up the original cuda variant.

$\endgroup$
1
$\begingroup$

It is difficult to make a broad statement about which hardware acceleration product has the best performance for deep learning until some comparative metrics using the same algorithms and data sets are published, which requires the availability of both products. It's anyone's guess from the below spec comparison. The information on the net is and could only be conjecture at this point.

    AMD Radeon VII
    Vega 20 GPU
    16 GB HBM2 memory
    1 Tb/s memory bandwidth
    60 compute units 
    3,840 GCN 1.8GHz stream processor cores 
    expected price $700 (possibly lower due to 7 nm lithography)
    software ecosystem set up difficulty: high

    NVidia RTX 2080 Ti
    GeForce RTX 2080 Ti GPU
    11 GB 352-Bit GDDR6
    14 GHz eff memory clock
    4,352 CUDA 1.5GHz stream processor cores
    PCIe 3.0
    current low side of newegg price $1,189
    software ecosystem set up difficulty: medium
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.