AMD Milan-X Upgrade In The Cloud Makes Microsoft Azure HBv3 Very Compelling For HPC
I had a few days to experiment with Milan-X on Microsoft Azure and it's been a real treat. Milan-X's incredible cache amount has opened up real performance benefits to a variety of workloads over an otherwise standard Zen 3 processor.
This initial benchmarking was done using an Azure standard_HB120-64r3_v3 (64 CPU core) instance and again using the standard_HB120rs_v3 (120 core) instance for getting a mix of coverage. Keep in mind this was single VM testing compared to some of the AMD and Microsoft numbers cited for benchmarking workloads across multiple VMs via MPI.
Both the existing HBv3 Milan and upgraded HBv3 Milan-X 64/120 core VMs were then tested using CentOS 8 and running a wide variety of workloads.
First up is the WRF weather forecasting open-source software as one of the areas cited by AMD as having enjoyed improvements with Milan-X. Indeed, the 3D V-Cache pans out well. Even with this being testing from a single VM rather than multi-VM deployment distributing the WRF workload, the improvement from having this huge CPU cache immediately shows itself to be beneficial... The Milan-X VMs shaved off around 10% of their run-time for the conus 2.5km model simply going from Milan to Milan-X. Now with Microsoft Azure making Milan-X available at the same cost as Milan, that's quite a nice mid-cycle upgrade.
Enjoying an even larger improvement going from Milan to Milan-X was the OpenFOAM computational fluid dynamics software. CFD is one of the areas AMD focused on with Milan-X and certainly the 3D V-Cache is paying off very well when dealing with very large workloads.
The benefits of Milan-X though are certainly contingent upon workloads with large enough data-sets. In some cases the performance uplift is much more minor, but still worthwhile especially as in cases like Azure where it's effectively a free upgrade over non-X Milan HBv3.
Even with our focus on open-source software rather than some of the commercial/proprietary software in the industry and focused in on for some of the AMD benchmark data out today, we are still seeing great improvements on these open-source packages.
Some of these benchmarks just scream for "more cache please!"