Radeon ROCm 3.8 Released With Hipfort For Fortran On GPUs, Data Center Tool

[phoronix]

Phoronix: Radeon ROCm 3.8 Released With Hipfort For Fortran On GPUs, Data Center Tool

Version 3.8 of ROCm, the Radeon Open eCosystem, is now available. This release continues making more progress on preparing the ROCm graphics compute stack for the upcoming large AMD supercomputer deployments and other data center usage...

http://www.phoronix.com/scan.php?page=news_item&px=Radeon-ROCm-3.8

[wizard69]

Has there been any hint from AMD about Navi support?

[Michael]

Has there been any hint from AMD about Navi support?

I've asked a few people about it, they said they would look into it, but have never gotten any response back to date.

[cb88]

Well apparently it's been working for about 2 weeks now.. https://github.com/RadeonOpenCompute/ROCm/issues/887

It isn't 100% stable but it is working. Perhaps some benchmarks are in order assuming you can get them to complete.

[oleid]

Well apparently it's been working for about 2 weeks now.. https://github.com/RadeonOpenCompute/ROCm/issues/887

It isn't 100% stable but it is working. Perhaps some benchmarks are in order assuming you can get them to complete.

This is for OpenCL, yes. But not for rocblas or MIOpen and thus tensorflow.

I got my R470 work with tf for simple networks but as soon as it gets a little more deep strange things happens and learning doesn't start (issue reported for MiOpen, but not heard back). But if it works out seems quite fast. Oh, and works with upstream kernel from Arch, which is nice. So I'm positive they will eventually get there, especially since they need it working for CDNA2.

[cb88]

This is for OpenCL, yes. But not for rocblas or MIOpen and thus tensorflow.

I got my R470 work with tf for simple networks but as soon as it gets a little more deep strange things happens and learning doesn't start (issue reported for MiOpen, but not heard back). But if it works out seems quite fast. Oh, and works with upstream kernel from Arch, which is nice. So I'm positive they will eventually get there, especially since they need it working for CDNA2.

You mean RX470... I wonder if the particular workload you are attempting is running out of vram?

In any case the rest will follow, and pytorch is being worked on also and should be close or working already. Also I agree, I expect CDNA to be quite similar to RDNA just optimized for compute.

[oleid]

You mean RX470... I wonder if the particular workload you are attempting is running out of vram?

Possibly, but it shouldn't. It's 1D data, about 90.000 parameters to train. Just lots of convolution. My guess would me memory fragmentation, but who knows.

Since the data is not public, I made a self contained minimal example with a simpler architecture yielding the same error. In case you're interested :

https://github.com/ROCmSoftwarePlatform/MIOpen/issues/436

[lumks]

does it work for RENOIR now or still freezes your whole PC on doing any CL workload (clinfo is enough)?

[cb88]

does it work for RENOIR now or still freezes your whole PC on doing any CL workload (clinfo is enough)?

That was apparently not a renoir specific bug but all the APUs.

https://github.com/RadeonOpenCompute/ROCm/issues/883

[oleid]

Didn't they say that 4000 series works?