Radeon ROCm 4.5 Released With HIP Runtime Compilation, Unified Memory Support

[phoronix]

Phoronix: Radeon ROCm 4.5 Released With HIP Runtime Compilation, Unified Memory Support

AMD today released Radeon Open eCosystem 4.5 (ROCm 4.5) as the latest version to their open-source GPU compute stack for Linux systems. ROCm 4.5 brings with it a number of new features and improvements but one area on the consumer Radeon side will leave some potential users frustrated...

https://www.phoronix.com/scan.php?page=news_item&px=Radeon-ROCm-4.5

[Sean_b2]

Disappointing as I thought rdna support (HIP) was required prior to Blender 3.0 release next month

[hoohoo]

Does AMD think it is a feature that almost nobody ca actually use ROCm? Radeon VII is no longer sold, mi50 is hard to find and expensive, mi60 is gone, mi100 afaict cannot even by purchased as a standalone product.

Great, ROCm is progressing nicely. Congrats. The ROCm team and the staff at US nat'l labs can have a party together.

Meanwhile the rest of the GPGPU world buys Teslas and RTXs and actually runs code.

Maybe ROCm is a better mousetrap but we'll never know, since we can't buy hardware to run it on

[Bengt0]

Used Vega 56/64 seem to be the way to go. They do not die early like the Radeon VIIs, 8 GB VRAM is still okay-ish and they are reasonably fast. PyTorch support is back up for them *just* now: https://github.com/RadeonOpenCompute/ROCm/issues/1572

[sliwka]

Looks like people have gotten ROCm 4.5 to work with the 6900XT:
reddit.com/r/hardware/comments/qkc6n8/6900xt_tensorflow_mlai_benchmarks_using_rocm/

Any chance we might get some benchmarks with it Michael ?

[Paradigm Shifter]

Does AMD think it is a feature that almost nobody ca[n] actually use ROCm?

Apparently so.

The frustrating element is still not having any support for "Navi" RDNA/RDNA2 graphics processors. AMD routinely has said they are working on enabling RDNA/RDNA2 graphics processors for ROCm but is clearly not a priority for them as they work on feature development for ROCm to compete with CUDA and ensuring their Instinct accelerators are primed for winning in the super-computing/HPC space. I have not seen any public indication when they intend to have RDNA(2) support in place for ROCm but unfortunate for those developers that may want to experiment with AMD's open-source compute stack but lacking the necessary hardware. This is an area where NVIDIA CUDA is in much stronger shape with their wide range of supported graphics processors.

Ne'er a truer word spoken. Or typed, in this case.

...

I appreciate the effort involved, and the difficulty of assigning limited resources, but this is fast becoming a bad joke. Every time I get my hopes up seeing a new ROCm release, my hopes are dashed just as quickly.

[ryuanlu]

At least we have opencl support for RDNA2.

[Bengt0]

If it is any indication, MIOpen supports RDNA2 since August: https://github.com/ROCmSoftwarePlatf...ses/tag/2.12.0

[Bengt0]

I have not seen any public indication when they intend to have RDNA(2) support in place for ROCm

Michael, my understanding is that RDNA(2) support will be in place for ROCm 5.0:

https://github.com/RadeonOpenCompute...ment-942101147

This should be right before the Christmas holiday season, or before the end of 2021, as stated earlier:

https://github.com/RadeonOpenCompute...ment-746164720

[satanas]

Does AMD think it is a feature that almost nobody ca actually use ROCm? Radeon VII is no longer sold, mi50 is hard to find and expensive, mi60 is gone, mi100 afaict cannot even by purchased as a standalone product.

So much of this, it isn't even funny! On top of that, no RDNA (2) support and even Vega, which is officially supported, is broken at places (see rocm 4.3 and pytorch on their github). Right now, there is no reasonable way to run ROCm as a consumer. This is not how you gain market share AMD, you're actually doing a perfect job going in the opposite direction..

If someone's listening, make ROCm support for consumer products a priority, right now!