With AVX-512 enabled and at the same power level, AMD EPYC 9004 ‘Genoa’ CPUs deliver a 35% performance boost.
The AMD EPYC 9654 processor is one of a new generation of server processors that were dubbed “the fastest server CPUs on the planet” upon their release, and Michael Larabel of Phoronix has put the new fourth-generation Genoa CPUs to the test in an impressive 130 benchmarks in the Ubuntu 22.10 OS environment.
In these benchmarks, no stone was left unturned — performance, temperature, frequency limits, and more were all tested to see how the new EPYC processor handles the recently implemented AVX-512 that was added to this new processor series.
The AVX-512 instruction set, first presented by Intel and incorporated into the company’s Intel Xeon Phi x200, Skylake-X, and recent Xeon Scalable processors, was introduced in AMD Zen 4. Each extension in the AVX-512 instruction set must be executed independently. AVX-512 has recently been used in several applications, including increasing performance. AMD claims that incorporating AVX-512 will improve performance and data management when processing video, analysing financial equations, and simulating scientific advancements.
Larabel has tested other AMD processors that support AVX-512, including the Ryzen 9 7950X and the EPYC 9004 series. In previous tests, AVX-512 was quite beneficial to both processors, demonstrating increased efficiency while keeping consumption and clock frequencies lower, particularly in large workloads. For his recent test, he used the AMD EPYC 9654 2P processor with AVX-512 active and deactivated in Ubuntu 22.10, which runs the current Linux kernel (v6.1).
In his artificial intelligence benchmarks, the performance with AVX-512 active was 35% higher (if not more in some cases) than with the instruction set deactivated. In AI workloads, processor power consumption was close to zero, but while active, the AVX-512 instances outperformed by lowering power consumption levels.
The Neural Magic DeepSparse 1.1 AI-related benchmark did show encouraging results for AVX-512 in the new AMD EPYC 9654 processor, but it was not as dramatic as some other machine learning workload tests. Neural Magic DeepSparse is a “sparsity-aware inference runtime” that provides graphics processing performance on processors and APIs, enabling machine learning integration.
AMD EPYC Genoa CPUs showcase a 35% performance improvement with AVX-512 enabled. (Image Credits: Phoronix)
Another AI-based benchmark, Mobile Neural Network 2.1, was a “odd duck” in the flurry of benchmarks because the AVX-512 implementation performed worse in only one test with the model “Inception-v3.” Larabel speculates that the software itself could be the catalyst but does not provide a definitive answer.
AMD EPYC Genoa CPUs retain the same power consumption even with AVX-512 enabled. (Image Credits: Phoronix)
Because crypto benchmarks and Tencent’s NCNN models were positive, the author switched to Intel-specific software that focused on the advantages of AVX-512. In AVX-512-enabled tests, AMD EPYC once again outperformed. Two instances of the Intel Open Image Denoise (v1.4.0) benchmark produced negligible results, but Larabel demonstrated that power consumption was still lower with active AVX-512.
Larabel concludes his testing for the time being, but notes that AMD’s Zen 4 architecture continues to outperform current Intel Xeon Scalable processors for this new generation, and it appears that even the upcoming Sapphire Rapids Xeon chips will struggle to compete against Genoa CPUs.
News Source: Phoronix
