AMD Preparing More Linux Code For The Frontier Supercomputer

[zxy_thf]

right but this is already a thing of the past... people run rocm opencl on RDNA 1/2 cards right now...
this problem was true for like 1-2 years for RDNA1 and like 6-8 months year for RDNA2...
i am on vega64... who has rocm support... so most sane people are not affected by this.
ROCm HIP is not yet done for RDNA1/2... but we will see it in the near future.

in my point of view AMD has a much better strategy than Nvidia... because of this they did make a lot of HPC sales against nvidia...

I ran ROCm OpenCL on my RDNA 1 card, but no luck on RDNA 2 card.
I'm guessing ROCm kernel driver needs some upgrades, but latest Fedora kernel is unsupported yet.

Anyway everything will be fine after a few months, and I'm not in a hurry to use OpenCL on my card for now .

[bridgman]

I ran ROCm OpenCL on my RDNA 1 card, but no luck on RDNA 2 card.
I'm guessing ROCm kernel driver needs some upgrades, but latest Fedora kernel is unsupported yet.

ok this is strange... bridgman told us that rocm opencl does work but the documentary do not tell us so..

What I said was that the ROCm components in the packaged amdgpu drivers support both RDNA1 and RDNA2, and were heavily tested on RDNA2.

I don't think we are testing Navi on the datacenter stack releases yet.

EDIT - hey, my test signature started showing up.

[zxy_thf]

What I said was that the ROCm components in the packaged amdgpu drivers support both RDNA1 and RDNA2, and were heavily tested on RDNA2.

I don't think we are testing Navi on the datacenter stack releases yet.

EDIT - hey, my test signature started showing up.

bridgman I found a problem in the latest 21.10 Linux driver. It claims support of 6700XT but its OpenCL doesn't work with that GPU.
After digging the source code I found 6700XT needs RoCM 4.2 for OpenCL and the problem was fixed after following instructions from https://rigtorp.se/notes/rocm/

Btw any plan to make RoCM available to other distros? Current supported platforms are not that great even when comparing with CUDA.