The AI Performance Benefit with AMX on Intel Xeon 6 "Granite Rapids"

2 hours ago 2

Besides the support for MRDIMM-8800 memory, another distinct advantage of Intel Xeon 6 "Granite Rapids" processors is the continued presence of Advanced Matrix Extensions (AMX). Here are some fresh benchmarks looking at the impact of AMX on the Intel Xeon 6980P processors for AI inference workloads.

Intel Xeon 6980P processor on heatsink

Back when Advanced Matrix Extensions debuted in 2023 with Sapphire Rapids processors, I ran some initial AMX performance benchmarks at the time. Last year when Granite Rapids launched with the Xeon 6980P processors, I had wanted to revisit the AMX performance comparison but then my AvenueCity reference server had failed prior to getting around to running said benchmarks. But now that I am back up and running with Granite Rapids using a Giga Computing R284-A92-AAL server, I recently carried out a fresh look at the AMX benchmarks.

Intel Xeon 6980P with AMX extensions

Today's article is looking at the dual Intel Xeon 6980P performance in various CPU AI workloads and then repeating the tests in the same hardware/software configuration but disabling AMX support. Disabling AMX support isn't available as a BIOS option or any easy Linux kernel boot argument. For software using AMX via the Intel oneDNN library the ONEDNN_MAX_CPU_ISA="AVX512_CORE_FP16" environment variable can be used to avoid the "AVX512_CORE_AMX" capability otherwise. The presence of AMX can also be concealed from user-space and kernel usage via the clearcpuid=598,600,601 kernel boot argument.

Intel Xeon 6980P with AMX soft-disabled

More than two years after AMX first began shipping on Intel Xeon processors, the selection of software able to make use of Advanced Matrix Extensions is still fairly limited but good coverage among popular AI packages. For this AMX on/off comparison the AI software evaluated included OpenVINO and Llama.cpp while running Ubuntu 25.10 with Linux 6.17 and using the two Xeon 6980P processors with 24 x 64GB MRDIMM-8800 memory.

Intel Xeon 6 Granite Rapids AMX Performance

When looking at the AMX default (enabled) versus disabled performance, the system power consumption, and CPU temperature were also monitored for seeing what if any impact AMX usage was causing on those vitals similar to AVX-512 back in the day. Thanks again to Giga Computing for supplying their R284-A92-AAL barebones server to make this new Xeon 6900P series testing possible at Phoronix.

Read Entire Article