4 ms·
hmm. 16 GB VRAM vs 3070 8 GB... looks better for fitting TF models in memory
by abledon 6y ago
hmm. 16 GB VRAM vs 3070 8 GB... looks better for fitting TF models in memory
- minimaxir 6y agoYou won't be fitting a TensorFlow model into an AMD GPU anytime soon. (despite many efforts to address that)
- abledon 6y agohttps://rocmdocs.amd.com/en/latest/ https://rocmdocs.amd.com/en/latest/ Isn't a solution?
- dogma1138 6y agoROCm is only available on Linux and doesn’t support NAVI GPUs. Overall if you need to have a solution that just works with any ML framework out there Radeon GPUs aren’t fit for purpose.
- FredFS456 6y agohttps://github.com/RadeonOpenCompute/ROCm#Hardware-and-Software-Support https://github.com/RadeonOpenCompute/ROCm#Hardware-and-Softw... Note that ROCm does not have support for even the current-generation of gaming GPUs (Navi).
- diab0lic 6y agoI have not tried this in a while but no. It requires a very involved setup compared to it's CUDA counterpart and certain functionalities don't work / performance characteristics for specific operations can vary wildly. If it look at the amount of time necessary to deal with it the calculus quickly swings in favor of just purchasing an Nvidia card. Edit: And some sibling comments have pointed out that it doesn't function on current gen cards. Nvidia still firmly leads the software game in this regard. I'm really hopeful this will change in the future but I had the same criticisms four years ago when I bought my last AMD card and nothing has changed yet.
- deeeeplearning 6y agoLet us know how that works out.
- trynumber9 6y agoThese cards are RDNA2, which is focused on gaming. If it happens to be good at any other workload that is incidental. Do not expect much support for this type of work on RDNA2. If you want compute, AMD will be announcing CDNA cards (perhaps Arcturus, which is allegedly 42TFLOPS in the same power envelope) sometime soon.