cancel
Showing results for 
Search instead for 
Did you mean: 

Infinity Hub Container Discussions

brewfalconenterprises
Journeyman III

Radeon RX6600XT, Ubuntu 22.04, and ROCm woes

I built a Linux machine (Ubuntu 22.04) specifically for the purpose of Deep Learning development and ROS2/Gazebo simulations. I selected a Ryzen CPU and Radeon (RX6600XT) GPU because I've never gone very deep with AMD products and wanted to try them out.

It took me several days of fiddling around with the amdgpu utility to get a seemingly workable configuration for ROCm. I tried building PyTorch and Tensorflow from source (4 days of effort) but consistently ran into brick walls that appear to stem from an unsupported GPU.

I finally gave in and attempted Dockers for PyTorch and Tensorflow. After noodling through the build.... the Docker still cannot access the GPU. Without GPU support, this machine is not fulfilling its intended purpose.

I'm really hoping AMD reps can provide suitable drivers for my OS/GPU combination, and point me toward the OEM-recommended build instructions. It was never this difficult with previous GPUs.

I truly want to be a believer in AMD. Please help me get there.

1 Solution
gc9
Adept III

More ROCm discussion can be found on github.com under 'RadeonOpenCompute', project 'ROCm', in the 'issues' and 'discussion' sections, such as issue 1698.  In summary, of RDNA2 cards, only the W6800 professional card is supported (tested, paid to fix) by AMD engineers, though some code can run on consumer/gaming gpu cards with similar gfx1030 gpu.  For gfx1031 gpu cards like RX 6700, or gfx1032 gpu cards like W6600 or RX 6600, some things can be made to work by setting an override environment variable 'export HSA_OVERRIDE_GFX_VERSION=10.3.0' before installing, so that ROCm uses gfx1030 code, but there are no guarantees.

Hardware and Software Support 

Support for RX6600XT (gfx1032,navi23)?

LLVM, AMDGPU Backend, Processors, GFX10.3 (RDNA2) 

RX 6700s for Machine Learning using ROCm 5.3

View solution in original post

8 Replies
gc9
Adept III

More ROCm discussion can be found on github.com under 'RadeonOpenCompute', project 'ROCm', in the 'issues' and 'discussion' sections, such as issue 1698.  In summary, of RDNA2 cards, only the W6800 professional card is supported (tested, paid to fix) by AMD engineers, though some code can run on consumer/gaming gpu cards with similar gfx1030 gpu.  For gfx1031 gpu cards like RX 6700, or gfx1032 gpu cards like W6600 or RX 6600, some things can be made to work by setting an override environment variable 'export HSA_OVERRIDE_GFX_VERSION=10.3.0' before installing, so that ROCm uses gfx1030 code, but there are no guarantees.

Hardware and Software Support 

Support for RX6600XT (gfx1032,navi23)?

LLVM, AMDGPU Backend, Processors, GFX10.3 (RDNA2) 

RX 6700s for Machine Learning using ROCm 5.3

I tried the HSA_OVERRIDE_GFX_VERSION trick previously to no avail. But, when I set the variable in the same command as running the script, it started working.

$ HSA_OVERRIDE_GFX_VERSION=10.3.0 python3.9 main.py

I got a very warm GPU to show for the effort. Thanks for the support!

0 Likes

The Radeon RX6600XT is a graphics card produced by AMD, and ROCm is a software platform for accelerating computing workloads on AMD GPUs. If you are experiencing issues with running ROCm on Ubuntu 22.04 with your Radeon RX6600XT, here are some possible solutions:

  1. Update your drivers: Make sure that you have the latest drivers installed for your graphics card. You can download the latest drivers from the AMD website.

  2. Check compatibility: Make sure that your version of ROCm is compatible with Ubuntu 22.04 and your graphics card. You can check the ROCm compatibility chart on the AMD website.

  3. Check your system configuration: Make sure that your system meets the minimum requirements for running ROCm. This includes having a supported version of Ubuntu, a supported version of the Linux kernel, and a supported version of OpenCL.

  4. Troubleshoot installation issues: If you are having trouble installing ROCm, try using the ROCm installation script or following the installation instructions provided by AMD.

  5. Check for conflicting software: Make sure that there are no conflicting software packages installed on your system that may be interfering with ROCm. This includes other graphics drivers, CUDA, or OpenCL drivers 8171.

  6. Check your system logs: If you are experiencing crashes or errors when running ROCm, check your system logs for any error messages or warnings. These can provide clues as to the source of the problem.

If none of these solutions work, you may need to contact AMD support for further assistance. They may be able to provide more specific guidance based on the specific issues you are facing.

0 Likes
devikaR
Journeyman III

Its a bit frustrating that there's no solution to this.

0 Likes

More than "a bit..." AMD really cannot support its own products.

0 Likes
n810
Journeyman III

Here's my install notes for the things I have working that may be helpful. [ Details will differ based on your video card, but these work for me ]  Guide includes some clues for where to find more info or do debugging.  Yes, it's different, but gets the basics working, you can go from there.
https://gist.github.com/nktice/b331615469edcb471587392721dec472

0 Likes
jmsilva
Journeyman III

Lady's and Gentlemen...

Have been working for a month on a project to use 2 AMD 7 7735HS(XFCE4 Desktops) WS/Servers with 1 AMD 5, 4 Raspberry PI4 machines working together as servers, all have Ubuntu Server 22.04 LTS installed, using Kubernetes to create a cluster of 1 master with 6 node slaves with virtualization of services, machines and websites under MaaS technology on the cloud using OpenStack and finally the Canonical Ubuntu Pro program which gives complete security against Cyber attacks and keeps all kernels updated without ever having to reboot a machine.

The purpose of this endeavor is to create not only automation, but also a bilingual cloud public service channel show casing all these technologies with news, health, hardware/software needed, etc…

The system is built all software installed, but the problems described on this thread need to be resolved if I am to be successful (multimedia software needs good drivers), so I am 100% invested on your hardware and their software. I also subscribed to the Ubuntu Pro 10 year server maintenance and security agreements which Canonical is promoting heavily, my point is, I got an interest/investment with both Canonical and AMD.

If you have any solutions I beg you let me test it, promise to be on time and work with you guys and the Canonical group as needed.

Thanks,

Juan Silva

jmsilva@notimundousa.com

NotiMundoUSA.com Bilingual Open Source news and commentary for We The People...
0 Likes
Smith00978
Journeyman III

I've been working on a project for the past month. The goal is to use two AMD 7 7735HS (XFCE4 Desktops) WS/Servers along with one AMD 5 and four Raspberry PI4 machines to function as servers. All of these machines have Ubuntu Server 22.04 LTS installed. We're using Kubernetes to create a cluster, consisting of one master and six node slaves. This cluster will handle virtualization of services, machines, and websites using MaaS technology on the cloud via OpenStack. Additionally, we're using the Canonical Ubuntu Pro program to ensure complete security against cyberattacks and automatic kernel updates without the need for machine reboots.

The main objective of this project is to establish an automated bilingual public service channel on the cloud. We plan to showcase various technologies, including news 8171, health information, and the required hardware and software.

We've completed the software installation, but there are some issues described in this thread that still need resolution, particularly related to multimedia software requiring good drivers. I'm fully committed to your hardware and its software. I've also subscribed to the Ubuntu Pro 10-year server maintenance and security agreements that Canonical is actively promoting. My point is, I have a vested interest in both Canonical and AMD.

If you have any solutions that could help, I kindly request the opportunity to test them. I assure you that I'll adhere to timelines and collaborate with your team and the Canonical group as necessary.

0 Likes