Google Cloud C3D Shows Great Performance With AMD EPYC Genoa

Back in August Google Cloud announced the C3D instances powered by AMD EPYC 9004 "Genoa" processors while only last week was C3D promoted to general availability. Curious about the performance of C3D after being impressed by AMD EPYC Genoa bare-metal server performance at Phoronix as well as what I've seen with Genoa in the cloud at Microsoft Azure and Amazon EC2 / AWS, here are some benchmarks of the new C3D up against other GCE instances.

In Google's announcement last week they noted the general purpose C3D VMs can offer up to 45% performance increases over prior generation VMs. Given what I've seen out of Genoa and the great Zen 4 benefits like AVX-512 and transitioning to DDR5 system memory, these claims aren't all that surprising. Some of the specifics Google shared in their announcement was around 54% better NGINX performance, MySQL and PostgreSQL database servers up to 62%, and in-memory databases like Redis up to 60%.

Google Cloud offers C3D VMs up to 360 vCPUs (the C3D vCPUs are made up of a mix of physical cores and SMT sibling threads) and up to 2.8TB of DDR5 memory. For the purposes of this initial Google Cloud C3D benchmarking on Phoronix, I was focusing on the 60 vCPU C3D VM instance. For this article the following instance types were compared:

c3d-standard-60: The new C3D AMD EPYC Genoa instance with 60 vCPUs, powered by AMD EPYC 9B14.

c2d-standard-56: The prior generation AMD EPYC Milan instance with EPYC 7B13 processor. This instance was at 56 vCPUs with not having a 60 vCPU option.

n2d-standard-64: The AMD EPYC Rome based instance with EPYC 7B12 processors. Here the closest sized instance was 64 vCPUs.

c2-standard-60: The Intel Xeon Cascade Lake competition at 60 vCPUs for reference.

Unfortunately for Google's new C3 machine types that are powered by the latest-generation Sapphire Rapids processors, they are sized at 22 / 44 / 88 vCPUs (among other smaller and larger sizes). So unfortunately nothing in the 56~64 range as tested with the other machine types, which is why there wasn't any C3 instance tested for this article due to the starkly different sizing. In any case this article is mostly focused on the AMD EPYC generational performance in Google Cloud.

All of the tested instances were on Ubuntu 22.04 LTS. In addition to raw performance the performance-per-dollar based on the current hourly pricing was also tabulated.