The AVX-512 Performance Advantage With AMD EPYC Bergamo

While this week was the surprise announcement of Intel AVX10 and with that taking the super-set of AVX-512 to both E and P core processors in the future, for next year's Xeon "Sierra Forest" server processors at up to 144 cores, it appears they will lack AVX-512/AVX10. Intel's AVX10 announcement noted initial support with Granite Rapids processors that will debut next year but no mention of the E-core-only Sierra Forest. With the AVX10 only coming to P/E core client processors after Granite Rapids, it would appear the high density Sierra Forest generation will miss out on AVX10/AVX-512 and not appear until Clearwater Forest. Meanwhile with the 128-core AMD EPYC 9754 "Bergamo" processors now shipping, there is AVX-512 with the Zen 4C cores. Here are some benchmarks looking at the AVX-512 impact for Bergamo.

Prior to learning about AVX10 I already had been working on some AVX-512 on/off benchmarks for Bergamo given that it's looking like Sierra Forest won't have AVX-512 (or now AVX10) and with Ampere Computing's AmpereOne processors there has also been no indication of Scalable Vector Extensions (SVE) support. So for now the 128-core Bergamo processors are in a rather unique position.

AVX-512 has proven to be very beneficial with Zen 4 for providing better AMD EPYC Genoa performance, efficient AVX-512 on the desktop, and even AVX-512 with Ryzen laptops and handhelds. With Bergamo I was curious about the AVX-512 impact being any different considering they are the smaller Zen 4C cores and quantifying that performance difference.

Today's article is looking at the AMD EPYC 9754 (1P) performance when running with AVX-512 and then repeating the benchmarks but with AVX-512 disabled. It's that simple and in addition to looking at the raw performance impact the CPU power consumption was monitored along with the CPU peak frequency and CPU temperature difference too.