Benchmarking Google's New "C2D" Compute-Optimized AMD Zen 3 Instances - N2D vs. C2 vs. C2D Performance
Yesterday Google Cloud launched their C2D instance type as their newest compute-optimized instances powered by AMD EPYC 7003 series processors. here are some quick benchmarks looking at the C2D performance against the existing Intel Xeon powered C2 instance type and the former EPYC 7002 based N2D instance type.
Google Cloud has already offered AMD EPYC 7003 / Zen 3 instance types such as with last year's Tau VM introduction and benchmarks. The new C2D instance type this week is simply Google's latest AMD cloud offering and is compute-optimized with a focus on CFD, EDA, and other high performance computing workloads.
The Google Cloud C2D instance family offers sizes up to 112 vCPUs backed by up to 56 physical CPU cores, up to 896GB of memory, and up to 3TB of local SSD storage. More details in the Google Cloud announcement.
Curious about the performance beyond the few numbers shared by Google in their announcement yesterday, I ran some benchmarks across the same size N2D / C2 / C2D instances. N2D was tested as the prior generation AMD EPYC 7002 series equivalent to the C2D line-up and also what Google used for comparing the generational performance in yesterday's blog post. The C2 instance type was also tested as the Intel compute-optimized instance alternative. While "C2", it's important to note that Google's latest Intel C2 instance is relying on Xeon Scalable "Cascade Lake" processors and not the latest Intel Xeon Scalable "Ice Lake" processors (Google currently doesn't have any Ice Lake based "Compute Optimized" instance type).
Due to minimizing costs for carrying out the comparison, just the "size 8" n2d-standard-8 / c2-standard-8 / c2d-standard-8 instance types were tested as the 8 vCPU configuration. Each instance type provided 8 vCPUs, 32GB of RAM, and were all tested using Google's default Debian 10 environment.
From testing in us-central1, the n2d-standard-8 instance costs $0.337968 per hour, the c2-standard-8 costs $0.4176 per hour, and the C2D is $0.3632 per hour. For at least the C2D instance tested, it was backed by an AMD EPYC 7B13 processor.
This is quite a straight-forward and quick introductory look at the Google C2D performance, so let's get straight to the performance and performance-per-dollar metrics looking at a few different HPC/compute focused workloads.