Latest Open-Source AMD Improvements Allowing For Better Llama.cpp AI Performance Against Windows 11

Written by Michael Larabel in Software on 17 September 2025 at 10:48 AM EDT. Page 2 of 5. 21 Comments.

First up is looking at the Vulkan performance out of Llama.cpp on the AMD Radeon RX 9070 XT.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Qwen3-8B-Q8_0, Test: Text Generation 128. Windows 11 25H2 was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Qwen3-8B-Q8_0, Test: Prompt Processing 512. Linux 6.17 + Mesa 25.3-dev was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Qwen3-8B-Q8_0, Test: Prompt Processing 1024. Linux 6.17 + Mesa 25.3-dev was the fastest.

For the text generation benchmarks the Vulkan performance out of Llama.cpp was quite comparable between platforms. But when it came to prompt processing the results were very interesting. Ubuntu 24.04.3 LTS with its stock HWE driver stack -- the same Linux 6.14 + Mesa 25.0 as shipped by Ubuntu 25.04 -- was much slower than Microsoft Windows 11 with the latest Radeon Software drivers. But now if springing ahead a few months to the in-development Linux 6.17 + Mesa 25.3-devel is game changing for the Llama.cpp Vulkan performance. In the prompt processing benchmarks the Vulkan performance with that leading-edge Linux software stack is now able to match or outperform the Windows results. This is a huge achievement for the open-source RADV driver stack.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: gpt-oss-20b-Q8_0, Test: Text Generation 128. Windows 11 25H2 was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 512. Windows 11 25H2 was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 1024. Windows 11 25H2 was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Text Generation 128. Linux 6.17 + Mesa 25.3-dev was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 512. Windows 11 25H2 was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 1024. Linux 6.17 + Mesa 25.3-dev was the fastest.

In some of the Vulkan benchmarks with different models the Linux performance still trailed Windows 11 with Radeon Software, but was at least no longer a huge disadvantage. This is wonderful seeing the Mesa RADV Vulkan open-source driver being able to perform so well against the Radeon Software stack on Windows. We're used to seeing fantastic RADV results for Linux gaming while it's great seeing RADV doing very well for AI too.

21 Comments - Next Page

Page: 1 2 3 4 5 Next Page

URL: https://www.phoronix.com/review/llama-cpp-windows-linux/2

⇱ Latest Open-Source AMD Improvements Allowing For Better Llama.cpp AI Performance Against Windows 11 - Phoronix

Latest Open-Source AMD Improvements Allowing For Better Llama.cpp AI Performance Against Windows 11