AMD EPYC Genoa CPUs show strong performance improvement with equally powerful AVX-512

AMD’s EPYC 9004 Genoa CPUs have been tested in several AVX-512 benchmarks by Phoronix and the latest Zen 4 parts seem to deliver a significant performance boost while retaining the same power.

AMD EPYC 9004 ‘Genoa’ CPUs offer a 35% performance boost with AVX-512 enabled and just as powerful

The AMD EPYC 9654 processor is one of a new series of server processors that was hailed as the “fastest server CPUs on the planet” at launch, and Michael Larabelle of Phoronix put the new fourth-generation Genoa CPUs to the test in an impressive 130 benchmarks in Ubuntu 22.10 operating system environment.

Nothing was left untouched in these benchmarks – performance, temperature, frequency limits and more were tested, to discover how the new EPYC processor handled the latest implementation of the AVX-512 added to this new processor series.

AMD Zen 4 introduced the AVX-512 instruction set, first introduced by Intel and integrated into the Intel Xeon Phi x200, Skylake-X, and recent Xeon Scalable processors. Each extension in the AVX-512 instruction set must be executed independently. Recently, the AVX-512 has been used in many cases, such as to increase performance. With AMD, the integration of the AVX-512 is reported to have improved performance and data management in video processing, financial equation analysis, and simulation in scientific developments.

Larabel has tested other AMD processors with an active AVX-512, such as the Ryzen 9 7950X and the EPYC 9004 series. In its previous tests, the AVX-512 was a perfect fit for both processors, showing increased efficiency while keeping consumption and clock frequencies lower, especially on higher workloads. . Used an AMD EPYC 9654 2P processor in its recent testing with the AVX-512 active and passive in Ubuntu 22.10, which uses the current Linux kernel (v6.1).

In its AI benchmarks, performance with the AVX-512 active was 35 percent higher (if not more in some cases) than the performance of the deactivated instruction set. Processor power consumption in AI workloads was very close to negligible, but while it was active, AVX-512 instances did a better job by keeping power consumption levels lower.

One set of AI-related benchmarks, Neural Magic DeepSparse 1.1, showed welcome results for the AVX-512 in the new AMD EPYC 9654 processor, but it wasn’t as dramatic as some of our other machine learning workload tests. Neural Magic DeepSparse is a “sparse inference runtime” that delivers graphics processing performance over processors and APIs, allowing machine learning integration. You can learn more about it here.

AMD EPYC Genoa CPUs show a 35% performance improvement with AVX-512 enabled. (Image credits: Phoronix)

Another AI-based benchmark, Mobile Neural Network 2.1, was the “odd duck” of a series of benchmarks where the AVX-512 performed worse, and it was only in one specific test with the “Inception-v3” model. Larabel states that it is possible that the program itself is the trigger but does not have a definitive answer.

AMD EPYC Genoa CPUs maintain the same power consumption even with AVX-512 enabled. (Image credits: Phoronix)

Tencent’s coding standards and NCNN models were favourable, so the author turned to Intel’s proprietary software focusing on the benefits of the AVX-512. Once again, AMD EPYC excels in the AVX-512-enabled tests. While running the Intel Open Image Denoise benchmark (v1.4.0), two cases showed negligible results, but Larabel showed that power consumption was still lower with the active AVX-512.

Larabel is finishing up its testing for now but notes that AMD’s Zen 4 architecture continues to show positive results for this new generation compared to current Intel Xeon Scalable processors and it appears that even the upcoming Sapphire Rapids Xeon chips will have a hard time competing with Genoa CPUs.

News source: Phoronix

Share this story



#AMD #EPYC #Genoa #CPUs #show #strong #performance #improvement #equally #powerful #AVX512

Leave a Reply

Your email address will not be published. Required fields are marked *