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

For the Nginx web server the Tau VM was able to match the performance of the Sapphire Rapids C3 VM while the C2D instance was performing similar to the Cascade Lake VMs. On a performance-per-dollar basis, the Tau T2D was leading.

Lastly was the BRL-CAD solid modelling open-source software where the C2D VM managed to nearly match the C3 VM while Tau was out in front.

From around 100 benchmarks carried out across these six different VM instance types in Google Cloud, below is the geometric mean based on all of the benchmarks:

Google's Tau T2D VM with 8 vCPUs easily dominated and came out 13% faster than the C3 Sapphire Rapids VM of the same size, but keep in mind Tau is special as Google backs each vCPU with a physical core rather than a mix of physical cores and SMT with their other VMs. That's quite a nice showing for the EPYC Tau VMs considering the t2d-standard-8 is one cent per hour less than the c3-highcpu-8 with the on demand hourly rate. The c2d-highcpu-8 instance using EPYC Milan processors and a 4 core + SMT mix like the c3-highcpu-8 managed to deliver 85% the performance of the Sapphire Rapids instance while having an hourly rate at 87% that of the Sapphire Rapids VM.

The C3 Sapphire Rapids VM was much faster than these generation-old EPYC VMs particularly for the workloads like AI that benefit from AVX-512 use. For workloads making use of all the latest and greatest CPU instruction set extensions, Sapphire Rapids was running great and much faster than EPYC Milan but for the rest of the time it was a rather tight race -- not just for Tau with its 8 physical cores but even the year-old C2D was still putting up a competitive fight against the new Sapphire Rapids VM. The race will be all the more competitive once Google Cloud rolls out any EPYC Genoa instances where there is now AVX-512 support, twelve channel DDR5 memory, etc. As shown since November in my many Genoa benchmarks, the latest AMD EPYC wares are extremely competitive in raw performance, performance per dollar, and also much better power efficiency although that is less of a concern for public cloud customers.

The EPYC Milan VMs performing so well against the new C3 Sapphire Rapids was a bit of a surprise after being so entrenched with Genoa the past few months, but in retrospect Milan and Sapphire Rapids were originally expected to be competitive partners in the market had Intel met their earlier targets of releasing Sapphire Rapids in 2021 and then 2022 but ultimately delayed to January 2023. So while Google's C3 Sapphire Rapids are a big upgrade over prior Intel instances when it comes to the EPYC Milan powered Tau T2D and C2D instances, overall it's still a competitive battle while it will be very interesting to see when Google introduces Genoa instances and how their pricing compares to C3.