ROCm 7.0.0 vs. ROCm 7.2.3 Performance On The AMD Radeon AI PRO R9700

Written by Michael Larabel in Display Drivers on 14 May 2026 at 10:20 AM EDT. Page 4 of 4. 4 Comments.

Blender benchmark with settings of Blend File: BMW27, Compute: Radeon HIP. ROCm 7.2.3 was the fastest.

Blender benchmark with settings of Blend File: Junkshop, Compute: Radeon HIP. ROCm 7.0.0 was the fastest.

Blender benchmark with settings of Blend File: Classroom, Compute: Radeon HIP. ROCm 7.2.3 was the fastest.

Blender benchmark with settings of Blend File: Fishy Cat, Compute: Radeon HIP. ROCm 7.0.0 was the fastest.

Blender benchmark with settings of Blend File: Barbershop, Compute: Radeon HIP. ROCm 7.2.3 was the fastest.

Blender benchmark with settings of Blend File: Pabellon Barcelona, Compute: Radeon HIP. ROCm 7.2.3 was the fastest.

When running Blender 5.1 on both ROCm 7 releases with the Radeon AI PRO R9700, there wasn't any real performance change observed with the Radeon HIP back-end across the variety of scenes tested.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: Qwen3.5-9B-Q8_0, Test: Text Generation 128. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: gpt-oss-20b-Q8_0, Test: Text Generation 128. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: Qwen3.5-9B-Q8_0, Test: Prompt Processing 512. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: Qwen3.5-9B-Q8_0, Test: Prompt Processing 2048. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 1024. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 2048. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: GLM-4.7-Flash-IQ4_XS, Test: Prompt Processing 512. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 2048. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: Mistral-7B-Instruct-v0.3-Q8_0, Test: Prompt Processing 512. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: granite-3.0-3b-a800m-instruct-Q8_0, Test: Text Generation 128. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: granite-3.0-3b-a800m-instruct-Q8_0, Test: Prompt Processing 512. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: granite-3.0-3b-a800m-instruct-Q8_0, Test: Prompt Processing 1024. ROCm 7.0.0 was the fastest.

Llama.cpp benchmark with settings of Backend: AMD ROCm HIP, Model: granite-3.0-3b-a800m-instruct-Q8_0, Test: Prompt Processing 2048. ROCm 7.0.0 was the fastest.

When adapting and rebuilding Llama.cpp for both ROCm releases, it was also a bit surprising to see no performance changes there from ROCm 7.0 to ROCm 7.2.3. Again if expanding the scope to include the AMDGPU/AMDKFD changes during that time would be the possibility of seeing more performance changes, but when it came to the ROCm user-space components there wasn't any real improvements observed for Llama.cpp and in fact a few regressions.

The most surprising takeaway from this testing was seeing some ROCm-OpenCL performance improvements. So that's the quick summary when comparing ROCm 7.0 to ROCm 7.2.3 stable... Now to look next at the latest TheRock-built ROCm Tech Preview.

4 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