Intel's Fantastic New Open-Source Demonstrator For AMX-BF16: Over 4x The Performance At 69% The Power

Written by Michael Larabel in Software on 13 January 2026 at 11:22 AM EST. Page 2 of 2. 38 Comments.

Once the first benchmark was complete, I was shocked... A 4x performance improvement from this release adding AMX-FP16 support for Granite Rapids CPUs. Freaking 4x improvement?!?! I ran it by Intel engineer Attila Afra who is a lead Open Image Denoise developer and he confirmed the performance improvement is aligned with their expectations. Wow.

But then it got even better as I was looking at the difference on CPU power consumption for this Xeon 6980P dual socket server too... The CPU power consumption was much lower too when running the OIDn 2.4 release. Dropping from a 500W average between the dual Xeon 6980P processors down to a 342 Watt average. Over 4x the performance and at around 69% the power consumption for this flagship Xeon 6 Granite Rapids AP server. Now that's a double win.

Running additional Open Image Denoise benchmarks continued to confirm the greater than 4x the performance from the updated Open Image Denoise leveraging AMX-FP16 on Granite Rapids. Prior to the 2.4 release, AVX-512 has long been the focus for capable Intel/AMD processors with this denoiser.

With the RTLightmap HDR 4096px test, the new Open Image Denoise 2.4 with AMX-FP16 support hit 4.6x the performance! While still consuming less power over OIDN 2.3 with just AVX-512 usage.

For the RTLightmap benchmark, Open Image Denoise 2.3 on the dual Xeon 6980P processors had a combined 634 Watts but dropped down to 435 Watts... Right around 69% the power while hitting 4.6x the performance with the software update adding AMX-FP16.

With Open Image Denoise used by Blender and numerous commercial applications (Arnold, V-RAY, Corona, Cinema 4D, RenderMan, Foundry Modo, etc) for high quality denoising, this AMX-FP16 support is quite a win for those rendering on Granite Rapids hardware. It's a great open-source demonstrator as well for being able to showcase the benefits of Advanced Matrix Extensions AMX-FP16 and something outside of conventional AI use with the big performance wins for AI workloads using AMX. It's just too bad that this software support only came more than one year after Xeon 6 Granite Rapids processors first began shipping.

For those after more Open Image Denoise hardware benchmarks, I continue to run OIDn performance benchmarks here. For those that haven't explored Open Image Denoise previously, you can find out more about this great open-source Intel software project at OpenImageDenoise.org.

38 Comments