AMD Ryzen AI Max "Strix Halo" Enjoys Great Performance Gains With Latest Linux Software

Written by Michael Larabel in Operating Systems on 2 April 2026 at 10:45 AM EDT. Page 4 of 4. 17 Comments.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: Qwen3-8B-Q8_0, Test: Prompt Processing 512. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: Qwen3-8B-Q8_0, Test: Prompt Processing 1024. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: Qwen3-8B-Q8_0, Test: Prompt Processing 2048. March 2026 was the fastest.

When firing up Llama.cpp with its Vulkan back-end, there were some especially solid performance gains out of the updated software stack provided by Ubuntu 26.04.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 512. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 1024. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 2048. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 512. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 2048. March 2026 was the fastest.

I had also wanted to test the Llama.cpp ROCm/HIP performance too, but the older Llama.cpp version wasn't building cleanly against the latest upstream ROCm packages. And wanting to keep the same Llama.cpp version to match what was originally tested. ROCm also isn't yet officially supporting Ubuntu 26.04 so there were a few workarounds needed there as well. In any event the Vulkan performance improvements with RADV on Strix Halo were quite impressive.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: Mistral-7B-Instruct-v0.3-Q8_0, Test: Prompt Processing 512. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: Mistral-7B-Instruct-v0.3-Q8_0, Test: Prompt Processing 2048. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: DeepSeek-R1-Distill-Llama-8B-Q8_0, Test: Prompt Processing 512. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: DeepSeek-R1-Distill-Llama-8B-Q8_0, Test: Prompt Processing 2048. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: granite-3.0-3b-a800m-instruct-Q8_0, Test: Prompt Processing 512. March 2026 was the fastest.

Llama.cpp benchmark with settings of Backend: Vulkan, Model: granite-3.0-3b-a800m-instruct-Q8_0, Test: Prompt Processing 2048. March 2026 was the fastest.

The Mesa RADV Vulkan driver performance has evolved very nicely over the past year not only for Strix Halo but Radeon graphics at large, especially RDNA3 and RDNA4 families.

Geometric Mean Of All Test Results benchmark with settings of Result Composite, AMD Strix Halo Linux Graphics March 2026 Ryzen AI Max. March 2026 was the fastest.

AMD Ryzen AI Max "Strix Halo" was impressive when it first launched and thanks to the ongoing open-source software improvements across the Linux stack, it continues aging very well. On a geo mean basis for the Radeon 8060S Graphics with the Ryzen AI Max+ 395, it was around 20% faster going from Ubuntu 25.04 to the near-final Ubuntu 26.04 packages. Benchmarks looking at the Ryzen AI Max CPU performance with Ubuntu 26.04 will be coming up in a separate article on Phoronix.

17 Comments

If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.

Page: 1 2 3 4