Radeon ROCm 5.4.1 Released

[phoronix]

Phoronix: Radeon ROCm 5.4.1 Released

AMD today capped off their busy week by releasing ROCm 5.4.1 as the newest version of their open-source compute stack...

https://www.phoronix.com/news/Radeon-ROCm-5.4.1

[Paradigm Shifter]

So at least for now there isn't official ROCm support for the Radeon RX 7900 series, but then again, AMD doesn't advertise the consumer cards with ROCm support... The GPU support in the documentation lists only the Radeon Instinct MI50 / MI60 / MI100 / MI200, Radeon VII, Radeon Pro VII, Radeon Pro W6800, and Radeon Pro V620 cards as being supported by ROCm. Much slimmer selection than NVIDIA's widespread range of CUDA support across their consumer and professional products.

Correct me if I'm wrong, but I'm pretty sure the Radeon VII was a consumer card. It was sure marketed like one, and available from places which do not traditionally sell workstation or HPC cards.

Irrespective of that... looks like I really should just give up on AMD GPUs completely and resign myself to my CUDA fate.

[skeevy420]

Correct me if I'm wrong, but I'm pretty sure the Radeon VII was a consumer card. It was sure marketed like one, and available from places which do not traditionally sell workstation or HPC cards.

Irrespective of that... looks like I really should just give up on AMD GPUs completely and resign myself to my CUDA fate.

The Radeon VII and the Radeon Pro VII might as well be the same thing outside of the Pro having better FP64 performance, ECC support, and no HDMI. Basically, the differences aren't enough to warrant excluding it like other consumer Vega GPUs.

[vegabook]

Can someone explain the technical reasons why we can't just use Vulkan compute for everything going forward? I mean, I get that Nvidia in particular has a bunch of domain-specific libraries that are really useful (Optix, Cudnn etc), but for vanilla compute, what's holding Vulkan back?

[Barley9432]

Can someone explain the technical reasons why we can't just use Vulkan compute for everything going forward? I mean, I get that Nvidia in particular has a bunch of domain-specific libraries that are really useful (Optix, Cudnn etc), but for vanilla compute, what's holding Vulkan back?

From what I've seen of the Blender effort, it's a VERY complex task to support Vulkan, and given just how amazing Optix and CUDA are, I imagine there isn't that much of a motivation.

Progress from the Blender side of things can be found here: https://developer.blender.org/T68990

https://devtalk.blender.org/t/blender-vulkan-status-report/25706

[oleid]

From what I've seen of the Blender effort, it's a VERY complex task to support Vulkan,

That's why some people have high hopes for compute in WGPU. I.e piet-gpu. But it's not finished, yet.

[NeoMorpheus]

Correct me if I'm wrong, but I'm pretty sure the Radeon VII was a consumer card. It was sure marketed like one, and available from places which do not traditionally sell workstation or HPC cards.

Irrespective of that... looks like I really should just give up on AMD GPUs completely and resign myself to my CUDA fate.

That I remember, amd was pretty much dead, thanks in great part to intel illegal tactics and some on nvidia part.

They didnt had anything to compete with nvidia on the consumer part and released the vega vii at a reduced priced, compared to the original server/pro version. Thats also why it was discontinued right away another consumer gpu was released by them ( i think it was the 5700xt) because it was really a very expensive gpu to be sold at the price they did.

[NeoMorpheus]

From what I've seen of the Blender effort, it's a VERY complex task to support Vulkan, and given just how amazing Optix and CUDA are, I imagine there isn't that much of a motivation.

Progress from the Blender side of things can be found here: https://developer.blender.org/T68990

https://devtalk.blender.org/t/blende...s-report/25706

Arent the blender group taking money from amd to make sure their crap runs on amd?

if thats the case, fuck the complexity and do the job that you were paid to do.

[mirmirmir]

Arent the blender group taking money from amd to make sure their crap runs on amd?

if thats the case, fuck the complexity and do the job that you were paid to do.

i don't think a company could tell blender foundation what to do. donation? sure they accept it. to support new hardware? sure, but your engineer should do the job.

[ripper81]

Arent the blender group taking money from amd to make sure their crap runs on amd?

if thats the case, fuck the complexity and do the job that you were paid to do.

This has lot to do with AMD'S focus when it comes to GPU computing. ROCm is only supported on Radeon Pro GPU'S not on consumer Radeon RX modells. The development is clearly focused on datacenters and partly professional content creation for beeing used with tools like Blender and Davinci Resolve.

While ROCm is opensource it seems to be difficult to package the thing for regular linuxdistribution. The installation and updates are not easy to handle even on the few officially supported distributions.

This is why projects like MESA rusticl are important. AMD doesn't seem to care about GPU computing on regular desktopsystems. The motivation to support linux as a desktop os is of course even lower than supporting windows.