AMD's ROCm 2.0 Radeon Compute Stack Being Prepared For Release

[phoronix]

Phoronix: AMD's ROCm 2.0 Radeon Compute Stack Being Prepared For Release

Last month AMD commented they would be releasing ROCm 2.0 prior to the end of 2018 and it looks like they will make good on their word. ROCm 2.0 is being prepared for release - source code is available albeit the reference Ubuntu/RHEL binaries are not yet out...

http://www.phoronix.com/scan.php?page=news_item&px=ROCm-2.0-Nearly-Released

[mannerov]

Does anyone know if ROCm/HSA has taken off somewhere ?

Around me, I see people using Cuda or OpenCL. No ROCm/HSA.

[Brisse]

N: Repository 'http://repo.radeon.com/rocm/apt/debian xenial InRelease' changed its 'Version' value from '1.9.2' to '2.0'

Received the update from AMD's binary repo just now

(and also Mesa 18.2.7)

[AndyChow]

Does anyone know if ROCm/HSA has taken off somewhere ?

Around me, I see people using Cuda or OpenCL. No ROCm/HSA.

You can now use TensorFlow, so it's pretty much out of the box. I would try it, but unfortunately I lack the pcie3, so can't try it. So if you use Tensorflow, then that's an option.

[acowley]

I hope AMD can get `hcc` and `hip` on a compatible release cadence with the rest of the ROCm stack. It's a nuisance that their releases pin different versions of `llvm` and `clang` than the tagged releases of AMD's forks, and don't seem entirely coordinated.

[marty1885]

Does anyone know if ROCm/HSA has taken off somewhere ?

Around me, I see people using Cuda or OpenCL. No ROCm/HSA.

Well OpenCL and CUDA/HIP on ROCm is built using... ROCm and HSA (Where ROCm is based on HSA). So......
Think like ROCm is like Mesa and OpenGL/Vulkan can be built on Mesa.

[Aeder]

You can now use TensorFlow, so it's pretty much out of the box. I would try it, but unfortunately I lack the pcie3, so can't try it. So if you use Tensorflow, then that's an option.

Is it upstream Tensorflow or are they still using their modified fork to make this work?

[nevion]

I've followed the project for a long time. I'm eager to see what 2.0 really means in time and how it will proceed since Greg has left, and I cannot gauge AMD's commitment very well. They often complain about too small a team which to me says amd, hire more people for this team - it's got prospects! Official RHEL support can only help them at least on the datacenter side where they are banking on cost effective machine learning servers and all the cloud related stuff.

As for who's using it, there's alot of interest in using it. There are several actual installations of it here and there and I've seen the hardware/software setups at universities, too. Cray seems to be looking strongly to them for their next round of servers, which should clue people in that there is value here. It's making rounds, and I can kind of half agree it's picking up steam but I also want to suggest it's not blooming out rapidly. I continue to have alot of hope that they'll eventually get the entire ecosystem leaps and bounds better than nvidia's distribution of cuda in terms of legit, fast compiler (clang), high performance, and supporting opencl again. I still have hopes also that opencl (2.x + c++ device side frontend) will be a real solution and not a limited scope strategy, but even if it doesn't I at least hope AMD shows NVidia how to manage the software technologies in this day and age. They're inching closer and closer to goodness while nvidia is innovating in other ways, not addressing the cuda "SDK"/nvcc all that much,

[enihcam]

If this is a modified tensorflow, AMD needs to provide a platform-independent binary. There are many rolling-release distro out there.

[ms178]

If this is a modified tensorflow, AMD needs to provide a platform-independent binary. There are many rolling-release distro out there.

The release notes only mention that PyTorch/Caffe2 still need a special ROCm fork.