AMD EPYC Milan Still Gives Intel Sapphire Rapids Tough Competition In The Cloud

Leela Chess Zero as an open-source chess engine making use of a deep neural network is an interesting test case here. Sapphire Rapids enjoys significant uplift over Cascade Lake in this benchmark, which makes use of AVX-512 for speeding up this AI chess engine. With the Tau t2d-standard-8 it isn't far behind the c3-highcpu-8 instance. The Tau T2D has much stronger performance than the c2d-highcpu-8 even though both are based on EPYC Milan because Tau is backed by eight physical cores versus 4 cores plus SMT for the other tested VMs. The c2d-highcpu-8 performance here was comparable just to Cascade Lake. The Tau showing was great while the EPYC Genoa performance should be even better since with Zen 4 there is the (efficient) AVX-512 implementation already talked about widely. Leela Chess Zero performs great with my Zen 4 hardware locally and should provide much uplift over these Google Cloud AMD numbers thanks to the AVX-512.

With the competitive pricing of the Tau T2D instances, the t2d-standard-8 on a performance-per-dollar basis comes in just behind the Sapphire Rapids C3 instance. On an hourly pricing basis, the t2d-standard-8 is about one cent per hour less than the c3-highcpu-8 VM.

With the miniBUDE HPC benchmark, the Tau t2d-standard-8 with its eight physical cores easily surpasses the c3-highcpu-8 Sapphire Rapids VM. Even more embarrassing, the c2d-highcpu-8 Milan instance matches the c3-highcpu-8 Sapphire Rapids performance!

With the strong AMD EPYC showing here, both AMD 8 vCPU instances tested deliver superior value over the tested Intel offerings.

In the case of NAMD and nekRS as two more HPC benchmarks, the Tau t2d-standard-8 instance delivered better performance than the Sapphire Rapids c3-highcpu-8 instance while costing less.