When it comes to software leveraging Intel Advanced Matrix Extensions (AMX) functionality in modern Xeon processors, it’s largely been limited to AI applications/libraries like oneDNN, OpenVINO, DeepRec, etc. But Intel now has another great open-source real-world AMX demonstrator with their Open Image Denoise library. This open-source library providing high quality denoising filters for images rendered using ray-tracing can end up benefiting big time from AMX-FP16 (AMX-COMPLEX) found with the newest Xeon 6 “Granite Rapids” processors. I ran some benchmarks of their new Open Image Denoise library with AMX-FP16 and was honestly blown away by the results.
Intel’s Open Image Denoise library is part of their (oneAPI) Rendering Toolkit for providing denoising for ray-traced images on both CPUs and GPUs. Intel Open Image Denoise in turn is used by software like the open-source Blender 3D modeling software as well as commercial applications like RenderMan, Cinema 4D, V-RAY, Autodesk Arnold and others for its powerful denoising capabilities.
Released last week was Intel Open Image Denoise 2.4 and with this new version comes support for Intel AMX-FP16 for “dramatically improving” the performance on capable processors. Do note though that this is explicitly AMX-FP16 “AMX-COMPLEX” for the half-precision floating point that to date is just found with Xeon 6 “Granite Rapids” processors. Xeon CPUs prior to Granite Rapids with AMX do not have the AMX-FP16 extension.
Curious about the AMX-FP16 performance called out for this non-traditional AI use-case plus with already using Open Image Denoising for various CPU/GPU benchmarks, I was eager to conduct my own independent benchmarks of the new release.
Using my currently only Intel Xeon 6 Granite Rapids server, the Gigabyte R284-A92-AAL1 with two of the flagship Xeon 6980P processors, I ran benchmarks of Open Image Denoise 2.3 compared to Open Image Denoise 2.4 with the AMX-FP16 support. I was using the default/reference binaries for each OIDn release.
No other changes were made to the Xeon 6 Granite Rapids Linux server during the testing besides swapping out the OIDn version being tested. Besides the AMX-FP16 support, the other Open Image Denoise 2.4 changes are focused on the GPU side. Open Image Denoise 2.4 adds support for Intel’s BMG-G31 GPU as well as the upcoming Wildcat Lake and Nova Lake processors with integrated graphics. Open Image Denoise 2.4 also supports the upcoming Crescent Island discrete graphics card even though that’s being marketed as an enterprise AI inference accelerator. Open Image Denoise 2.4 also improves non-Intel GPU support too with now having AMD RDNA 3.5 GPU support, extended AMD RDNA 2 support, support for building against NVIDIA CUDA 13, and support for building against AMD ROCm 7.
But let’s see how this AMX-FP16 introduction to Open Image Denoise affects the performance and power…
