AMD released a tool that converts CUDA code to C++ which can run on both Nvidia and AMD GPUs.

Vivek Suvarna Dec 15, 2016

  1. This is a big deal, especially for software which insists on Nvidia cards for CUDA support. This tool allows developers to convert their CUDA code to portable C++, which can be complied to run on both Nvidia and AMD graphic cards.

    Here's the link: https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP
     
  2. Very interesting. Thanks for sharing.
     
  3. The advantage with CUDA is of course performance. Usually these converted code solutions lag in terms of performance, but this might be convenient way for devs to build for CUDA but still support other GPUs.
     
  4. In comparison to what? AMD cards are also not bad and NVidia seems to "slow down" open approaches like OpenCL to force developers to CUDA and make the software Nvidia-dependent.

    As I'm doing some development and all my cards are Nvidia, I'm also mostly using CUDA althought I'd prefer to have more portable code so I'll have a look at HIP. Looks interesting.
     
  5. Ingmar is right. FCPX for example, is one of the softwares that relies on OpenCL and AMD cards seem to perform better with it.

    Its not that one is better than the other. Both have their strongsuits.

    Though I have heard a few complaints about how OpenCL documentation is maintained. If the devs aren't happy with it, it won't last long and that could be a cause for concern.
     
  6. It's not just the performance difference (which is quite considerable). OpenCL is a pain in the hole to write and compile in comparison to CUDA. Anything that eases the process of translating it to OpenCL is very welcome.
     

Share This Page