Radeon ROCm 3.5 Released With New Features But Still No Navi Support

[bridgman]

I still have a number of GCN1.0 gpus on hand including R9 280X, which should be very good at least at OpenCL 1.x, but it is not supported and would not be supported by ROCm.

IIRC the 280X was right on the edge in terms of being supportable - it had many of the HW features required for ROCm (specifically the MEC engines with microcode cache rather than fixed-size store) but because the HW shipped before the ROCm stack was developed it wasn't *quite* there. Might be possible to make it work if you have a lot of them though...

[_r00t-]

Strange, are you sure the card is okay?

https://gitlab.freedesktop.org/drm/amd/-/issues/914 it's a driver issue...

https://bugzilla.kernel.org/

They already have report this issue

[birdie]

https://bugzilla.kernel.org/

Half of the bug reports filed against AMD GPUs have zero replies from AMD developers/employees. Nice try! Nice AMD!

[zxy_thf]

I think the best solution for AMD consumers would be the MESA OpenCL implementation. Now that it seems it is picking up support, it might become viable soon. The majority of people just need OpenCL support, they don't care about the more advanced features ROCm brings. I personally just want to be able to run some OpenCL stuff without having to rely on the binary module.

Well it depends. For blender users it would be okay, but for ML users OpenCL is poorly supported by Tensorflow and not supported by PyTorch at all.
For GPU developers, my personal experience is OpenCL is really hard to work with and I would like alternative solutions.

[Marc Driftmeyer]

https://en.wikipedia.org/wiki/List_o...L_applications

The app that no longer uses OpenCL is Final Cut Pro as it now uses Metal, along with Logic Pro X and many other 3rd party OEMs Audio Unit Plugins, etc.

https://www.apple.com/newsroom/2019/10/final-cut-pro-x-update-introduces-new-metal-engine-for-increased-performance/

https://en.wikipedia.org/wiki/Metal_(API)

[CochainComplex]

...now a full year later I have to download something and compile it? And then run make install and wreak havoc to my system? .

Indeed this is how you install the Nvidia driver.

[paupav]

Still waiting for southern islands support/s

[bug77]

Indeed this is how you install the Nvidia driver.

Download source code and compile it? You must be kidding, I haven't had to do that for an Nvidia driver in years.
(I hope you were not hinting at dkms, that would be lame)

[coder]

run Nvidia if you have anything related to GPU Compute! Its not one click setup, but it works and thats probably the key feature

That's not strictly true. They don't support CUDA on some very low-end cards, like GTX 1030.

As far as "anything related to GPU Compute", most of their GPUs top out at OpenCL 1.2, but their Jetson kits don't support it at all.

[StillStuckOnSI]

IIRC the 280X was right on the edge in terms of being supportable - it had many of the HW features required for ROCm (specifically the MEC engines with microcode cache rather than fixed-size store) but because the HW shipped before the ROCm stack was developed it wasn't *quite* there. Might be possible to make it work if you have a lot of them though...

As you can tell, I have a vested interest in this :P

Would "might be possible" be as simple as installing the correct ROCm libraries/runtimes and crossing one's fingers? Or is it more a matter of upstream adding some special-cased code paths for certain GCN 1.0 devices if enough folks can successfully advocate for such a thing?