Replies: 27 comments 17 replies
-
7900xtx and similars including laptops are working now, but not has official support. For the whole community it is already great news, they have finally realized that the whole opensource community (consumer gpus) and PhD students can greatly accelerate the progress of this library. Hopefully they will also give native windows support soon, although it is optional, almost everyone uses linux for AI. |
Beta Was this translation helpful? Give feedback.
-
|
What about Vega56 / GFX9 is it still supported somehow? |
Beta Was this translation helpful? Give feedback.
-
|
While the development of NVK (a new open source Nvidia driver) is advancing, Nvidia is supporting their GPUs for at least 8 years, and already have the great (and mature) CUDA, I really can't see why a business or an individual should take a risk and go with AMD. I want to support AMD, I really do. But the reasons above are hard to ignore. Edit: It turns out NVK is only for Vulkan but it's an improvement still. We'll get there eventually with these. |
Beta Was this translation helpful? Give feedback.
-
|
as an RDNA user and I'm starting to be interested in AI there is no way to go with new AND GPU just because you don't went support old GPU like my RX 5700 XT CUDA looks so sexy from here when I'm trying experiment with Stable Diffusion I moved from windows and linux just to use hack environment variable "HSA_OVERRIDE_GFX_VERSION=10.3.0" to make Stable Diffusion work |
Beta Was this translation helpful? Give feedback.
-
|
As a long time (10+ years) CUDA developer, I just bought a Radeon VII (because it was one of the few cards AMD appeared to actually support) so I could try out developing for AMD hardware, too. It's also a great card in the sense that it's one of the few options on the market (from any vendor) to have good FP64 performance at a reasonable price point. It's disappointing to see AMD orphaning capable hardware like this so early in its lifetime. |
Beta Was this translation helpful? Give feedback.
-
|
As Radeon VII user and since 2005 using starting with Ati area bought always AMD but this thing force me why I am still going with AMD which I can't get support and suffer so much... |
Beta Was this translation helpful? Give feedback.
-
|
I have bought recently 3 Radeon VII for some projects at work (which is... AMD). So this is sad to have them no longer supported soon. :-( |
Beta Was this translation helpful? Give feedback.
-
|
Also, it seems like AMD team has some confusing intentions because AMDKFD interface in the Linux kernel still supports all cards started from GCN, and effectively only ROCm uses such an interface. (rusicls also thinking about it but not yet) I am a Vega56 user. |
Beta Was this translation helpful? Give feedback.
-
ROCm is open source but not open development. They do even not read the issue tracker usually. |
Beta Was this translation helpful? Give feedback.
-
|
Yes, please keep supporting gfx906 on ROCM for another few years at least! I use a Radeon VII for HIP training courses and it has by far the best double precision compute available in a consumer grade GPU. It is environmentally irresponsible to force users to discard hardware that still has plenty of life left! |
Beta Was this translation helpful? Give feedback.
-
|
+1 gfx906 support I waited for years to see ROCm come to Windows, only to find out that my Radeon VII, which was considered one of the best consumer GPUs only 2 years ago, wouldn't be included, not even in future releases! Please consider extending hardware support 🙏 |
Beta Was this translation helpful? Give feedback.
-
Nvidia still supports eight year old 2015 GPUs. AMD is dropping support for four year old 2019 GPUs. When we finally thought we had a feasible Nvidia alternative, it was too good to be true. Can’t believe we have to rely on Intel of all companies to offer a sustainable alternative. |
Beta Was this translation helpful? Give feedback.
-
|
Because Nvidia not change cuda cores drastically over years. AMD is really new with cdna and rdna so in my opinion they are working on these architectures right now, gcn is totally different. the problem was the strategy of amd with raytracing and AI, not giving the sufficient importance. Now, nvidia is moving to path tracing and tensor cores and rt cores are quite fast to run algorithms and DL. Rdna3 is not sufficiently strong in these areas but with rdna4 i’m pretty sure that they put AI hardware |
Beta Was this translation helpful? Give feedback.
-
|
Added support for additional GPU architectures. |
Beta Was this translation helpful? Give feedback.
-
|
For me it's working very well. |
Beta Was this translation helpful? Give feedback.
-
|
Not sure if this should be added as another issue or placed here... As pointed out by @stalkerg, rocm is "open source but not open development.". Where decisions to drop support for hardware is made behind closed doors with no input from normal users or clear reasons why. It appears there is no technical reason why 5 year old AMD cards cannot be supported for a much longer period. I speculate that the reason for AMD's decision is data-centers and HPC customers don't use cards older then 4-5 years... As others have pointed out, it is environmentally irresponsible to cut support or usability of a perfectly capable piece of kit. I, like most people, hope AMD do the right thing. Community/AMD, is it possible to (re)add support for formally dropped cards? From @stalkerg's comment the limitation is not in the driver (amdgpu), but within rocm itself (?). Is there any documentation or information about adding/maintaining support in rocm for unsupported GPUs? Would AMD be open to accepting PR's from the community to support unofficial/dropped cards? |
Beta Was this translation helpful? Give feedback.
-
Hi :) Thanks =) |
Beta Was this translation helpful? Give feedback.
-
|
I want to buy an Radeon VII pro, the specs are great and the price is great, but this is legitimately insane. |
Beta Was this translation helpful? Give feedback.
-
I totally agree, but I have to correct you about the age of the no more supported Kepler GPUs mentioned for comparison: actually they are 2012 GPUs! |
Beta Was this translation helpful? Give feedback.
-
|
@ppanchad-amd do you have any valid comment here? Is it a official answer? |
Beta Was this translation helpful? Give feedback.
-
|
Reposting an opinion from Phoronix that quite well summarizes my stance on hardware support on ROCm.
Unless AMD addresses those issues seriously I'd rather root for any other alternative that be NVIDIA opensourcing their driver stack, Intel pushing oneAPI, Vulkan compute or anything else that can give us working compute stack on Linux. Even thou I have AMD GPU in my own Linux machine, when it came to advising purchases for an institution that I work for I advocated for NVIDIA specifically for the reason explained above. |
Beta Was this translation helpful? Give feedback.
-
|
There seems to be some effort going on about supporting SPIR-V based kernels of ROCm: llvm/llvm-project#89796. Speaking from experience developing ROCm projects, these things usually take quite some time, and I expect (or rather hope) that this SPIR-V supports lands in the rest of the ROCm stack sooner or later. This would enable ROCm projects to support a wide array of different GPUs more easily. This the equivalent of PTX that the phoronix comment points out. On a side note, HSA (Linux ROCm's backing driver interface) already supported a PTX equivalent in the past, HSAIL. I guess ROCm never adopted it... |
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
-
|
I just saw this thread and to put it simple... The entire mid-range and low-end unsupported (laptops with 680m/780m integrated, better not talk about those with VEGA graphics) and drivers that break the performance of ROCm. |
Beta Was this translation helpful? Give feedback.
-
|
I just saw... |
Beta Was this translation helpful? Give feedback.
-
|
Shoutout to why re-instating MI50 support would simply be the right thing to do: |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
I was utterly amazed to read this in the ROCm 5.6 release notes:
The Vega 20 GPU (gfx906) is barely 5 years old. Even more, the Radeon VII, Radeon Pro VII and Instinct MI50 are still being sold!
If you want to know why everyone is going with Nvidia, there are a lot of them, but what they do at least well is they support they GPUs.
In the CUDA 12.x release from December 2022, Nvidia also dropped support for some GPUs:
Yes, you read that right, Kepler GPUs. These are 2014 products. Not 2019 products like you're dropping.
So, hereby I would you strongly urge to not force us all to Nvidia again, by properly supporting your products.
Beta Was this translation helpful? Give feedback.
All reactions