AMD EPYC Milan Still Gives Intel Sapphire Rapids Tough Competition In The Cloud
While waiting for AMD 4th Gen EPYC "Genoa" instances to become available via the major public cloud providers, I was curious to see how existing AMD EPYC Milan instances compare to Intel's new Sapphire Rapids instances in public preview on Google Cloud. While expecting some friendly competition, at the same vCPU size EPYC Milan was managing to deliver not only better performance-per-dollar but also even better raw performance in numerous workloads against the Google Cloud C3 Sapphire Rapids.
With my recent tests looking at the Google Cloud C3 "Sapphire Rapids" performance for those new Intel 4th Gen Xeon Scalable VMs available via Google Cloud, the Sapphire Rapids VM at the same VM size was delivering dramatic uplift over the still-very-common Intel Cascade Lake VMs. While no EPYC Genoa VMs are publicly available yet with Google Cloud, curiosity got the best of me to see how the Sapphire Rapids C3 VMs performed to 3rd Gen EPYC.
All the VMs tested were the "-8" (8 vCPU) size given that no longer C3 Sapphire Rapids instances were available during my testing. For all except Tau this 8 vCPU configuration for Intel/AMD means four physical CPU cores plus SMT/HT to deliver 8 vCPUs. With Google's Tau T2D flagship VMs, they are backed by eight physical cores without SMT being active.
The configurations for this round of testing were:
- c2-standard-8 (Cascade Lake)
- n2-standard-8 (Cascade Lake)
- n2-highcpu-8 (Cascade Lake)
- c3-highcpu-8 (Sapphire Rapids)
- c2d-highcpu-8 (Zen 3 / Milan)
- t2d-standard-8 (Zen 3 / Milan)
As outlined in the prior article, Google isn't running any prior generation Ice Lake VMs at the -8 size. Presumably once Google introduces the EPYC 9004 "Genoa" series, they will be the basis of the C3D series to complement the C3 Sapphire Rapids offerings.
For all VMs tested they were backed by a 300GB SSD disk and running Ubuntu 22.10 for running an up-to-date Linux stack with the current GCC 12 stable series, Linux 5.19 kernel, and other fresh software components. From the software side and benchmarks everything was the same with just switching out the Google Cloud VM/instance type.
In addition to looking at the raw performance, the performance-per-dollar was also analyzed using the hourly pricing.