AMD EPYC Milan-X CPUs and Microsoft Azure HBv3 VMs appear to be a perfect fit for cloud-based server usage, showing staggering performance upgrades
Microsoft is not planning to change the pricing of the new HBv3 at this time, especially when they are handling the changes as an upgrade and not a complete suite offering. Microsoft realizes that Milan-X is exceptionally influential for large HPC workloads deployed across multiple VMs. However, Microsoft Azure is now handling higher competitiveness from Google Cloud Tau VM and other cloud providers. The significant advantage is the price point does not increase with the newest integration of Milan-X. Consumer benefit for the newest update allows users to integrate code into AMD 3D V-Cache for future tweaking with the ability to utilize such an increased cache size and breathing room for testing workloads using the updated system cache. Larabel used the previous benchmark results from November 2021 as a baseline for the new testing of the HBv3 VMs. The first benchmarking was accomplished using an Azure standard_HB120-64r3_v3 (64 CPU core) instance and the standard_HB120rs_v3 (120 core) instance to see the varying differences between the two cores. Larabel used the same two cores for the updated benchmark tests. Larabel’s VM testing was singular compared to AMD and Microsoft, citing numerous benchmarking workloads across multiple VMs via MPI. The standard HBv3 Milan and boosted HBv3 Milan-X 64/120 core VMs were analyzed using CentOS 8 and diverse workloads. It is to note that Milan-X is essentially AMD Milan with their 3D V-Cache added to current Zen 3 cores. Highlights of Larabel’s retesting are:
AMD’s increased cache amount on Milan-X shows significant performance advantages to mixed workloads over the standard Zen 3 processors. AMD touted WRF weather forecasting open-source software to receive valued performance under the new Milan-X. The new chips reduced 10% on run-time for the CONUS 2.5km model. OpenFOAM computational fluid dynamics (CFD) software also revealed increased performance from the upgrade. AMD focused heavily on CFD with the company’s Milan-X. The benefits of Milan-X, though, are undoubtedly contingent upon workloads with large enough data sets. Performance increases were minor for data set workloads, but it is still beneficial to switch with the no-cost upgrade. Open-source packages also saw improvement in benchmarks. Very few tests revealed that more cache was necessary. Milan-X showed valuable improvement in other HPC workloads not backed by AMD or Microsoft. Zstd, with the settings set to the highest compression, gained adequate results from the AMD 3D V-Cache. Specific imaging workloads show significant progress. Code compilation speed was minor.
AMD Milan has a significant lead versus Intel’s Xeon Scalable Ice Lake. With the benchmarks completed by Larabel, the new Milan-X pushes further ahead with many of the workloads performed. AMD EPYC 7004 “Genoa” chips are set to launch during the second half of 2022. Source: Phoronix
