Arch Linux User Repository

I have created four new packages for the ROCm technology preview release stream - currently ROCm 7.10.0. I assume with time they will be able (and forced) to replace the current opencl-amd-dev PKGBUILD.

• rocm-gfx120x-bin for RDNA 4
• rocm-gfx110x-bin for RDNA 3
• rocm-gfx1150-bin for Ryzen AI
• rocm-gfx1151-bin for Ryzen AI Max

Latest release: 7.2.0. It uses 16.45GB of disk.

Issues with current release:

• libxml2-legacy is required by LLVM compiler.
• libamdpython.so is missing. Needs manual intervention for rocgdb to function.
• ROCm needs OpenCV libraries. Possible solution is install opencv then create symbolic links to the Arch Linux libraries.

rocm-gfx120x-bin 7.11.0pre-1 working fine for me on arch based distro ( cachyos )

Thanks you very much @luciddream for your hard work. :)

Just commenting here about the other 4 (and soon 5 packages). There seems to be an issue with the build process of the 7.11 preview release, confirmed by the AMD team, which probably affects the library search order. I will make more tests tomorrow because it's almost 01:00 am here. We can probably bypass the issues with setting some paths here and there

Hi all, please check it's working for you, I was too tired when I made the release.

I have created four new packages for the ROCm technology preview release stream - currently ROCm 7.10.0. I assume with time they will be able (and forced) to replace the current opencl-amd-dev PKGBUILD.

• rocm-gfx120x-bin for RDNA 4
• rocm-gfx110x-bin for RDNA 3
• rocm-gfx1150-bin for Ryzen AI
• rocm-gfx1151-bin for Ryzen AI Max

@luciddream Yes, it’s a test I made to compare and test the performance of OpenCL. As soon as I solve some issues, I’ll publish it for the community. :)

@cmhacks cool, is that benchmark something you have made?

@luciddream All working like a charm. Many thanks for your hard work dude :)

❯ ./rocm-9070xt-test

================================================================================
                       ULTIMATE GPU BENCHMARK SUITE
================================================================================
ROCm 7.1.1 Optimizations Applied:
  • HSA_OVERRIDE_GFX_VERSION=12.0.1 (RDNA4 Fix)
  • ROCM_FORCE_WAVE32=1 (MFMA Boost)
  • HIP_ENABLE_GRAPHS=1 (Lower Overhead)

TIP: For maximum stability/performance, run these commands manually:
  echo "manual" | sudo tee /sys/class/drm/card0/device/power_dpm_force_performance_level
  echo "5" | sudo tee /sys/class/drm/card0/device/pp_power_profile_mode
--------------------------------------------------------------------------------
Detected ICDs:
  • ROCm: /etc/OpenCL/vendors/AA_amdocl64.icd
  • Pocl: /etc/OpenCL/vendors/pocl.icd
  • Rusticl: /etc/OpenCL/vendors/ZZ_rusticl.icd
--------------------------------------------------------------------------------

--- Detected OpenCL Devices ---
[a] Run All (ROCm + Rusticl + HIP + Matrix) - Default
Select device [a]:

Running benchmarks with warmup and sustained load (min 1s)... Please wait.


=== BENCHMARK EXECUTION ===
Running ROCm (OpenCL)...                Done
Running Pocl (OpenCL)...                Done
Running Rusticl (OpenCL)...             Done
Running HIP Vector (Shaders)...         Done
Running Matrix Cores (AI)...            Done

=== PERFORMANCE REPORT ===
+----------------------+----------------+----------------+----------------+----------------+----------------+
| METRIC               | ROCm           | Pocl           | Rusticl        | HIP            | MATRIX (AI)    |
+----------------------+----------------+----------------+----------------+----------------+----------------+
| FP64 (TFLOPS)        | 0.60           | 0.05           | 0.00           | -              | -              |
| FP32 (TFLOPS)        | 4.68           | 0.34           | 4.66           | 4.77           | -              |
| FP16 (TFLOPS)        | 0.54           | 0.05           | 0.57           | -              | 164.00         |
| BF16 (TFLOPS)        | -              | -              | -              | 0.60           | 161.34         |
| INT64 (TOPS)         | 0.08           | 0.01           | 0.07           | 0.08           | -              |
| INT32 (TOPS)         | 0.15           | 0.01           | 0.15           | 0.15           | -              |
| INT16 (TOPS)         | 0.29           | 0.03           | 0.28           | 0.28           | -              |
| INT8 (TOPS)          | 0.44           | 0.07           | 0.53           | 0.45           | 86.68          |
| FP8 (TFLOPS)         | -              | -              | -              | 0.58           | 83.81          |
| INT4 (TOPS)          | 1.20           | 0.10           | 1.18           | 1.20           | 0.00           |
| FP4 (TFLOPS)         | 1.21           | 0.10           | 1.18           | 1.21           | 0.00           |
| FP6 (TFLOPS)         | -              | -              | -              | -              | 0.00           |
+----------------------+----------------+----------------+----------------+----------------+----------------+

Note: All values in TFLOPS/TOPS. 'Matrix' uses hipBLASLt.
Benchmark Complete.

::BENCHMARK_COMPLETE::

=== SYSTEM RECOMMENDATION ===
Based on the benchmark results:
  • ROCm Wins: 7
  • Pocl Wins: 0
  • Rusticl Wins: 2

Recommendation:
  ✅ Use ROCm as your default OpenCL driver.
  It performed better in 7 tests.
::WINNER::ROCm

(Note: For AI/ML workloads, always use HIP/ROCm regardless of OpenCL choice)

Report saved to: ./reports/rocm_benchmark_report_1764369720.txt

hey @cmhacks, I just made a new release. It's very rushed but hopefully I didn't forget anything

Good day @luciddream

Eta for 7.1.1 ?

Thank you very much for your hard work! :)

@luciddream I just got lost in the infinite web of the everchanging dependencies on ROCm right now, your observation is on point.

Let me try to get time and mental bandwidth to try something better for my specific scenarios.

Thanks anyway!