Azure Provides Excellent HPC Cloud Performance With HBv4 Series Powered By AMD EPYC Genoa-X

The past several weeks at Phoronix has been very exciting with benchmarking AMD EPYC Genoa-X processors (along with Bergamo) and the incredible uplift delivered by these latest AMD server processors. But for those not yet having the opportunity to test the new EPYC Genoa-X processors locally, those wishing to evaluate the Genoa-X capabilities in the public cloud prior to making an investment in these high-end server processors with 3D V-Cache, or those simply preferring the ease of cloud infrastructure, Azure's new HBv4 series provide an excellent route for leveraging AMD Genoa-X compute capabilities in the cloud. Here are benchmarks of the new Azure HBv4 powered by EPYC Genoa-X compared to prior Azure HPC VMs. The Azure HBv4 performance is outstanding with incredible generational uplift and leading value among Microsoft's HPC-focused VMs.

Microsoft Azure's HB-series VMs have been tailored for workloads needing high memory bandwidth from FEM to CFD and other heavy computational workloads. Last year Microsoft upgraded the Azure HBv3 series with Milan-X and since that first generation 3D V-Cache server processor deployment, these cloud VMs have offered industry-leading performance. Now with the HBv4 series the performance is seriously super-charged thanks to Genoa-X. With Genoa-X's 3D V-Cache, AVX-512 with Zen 4, DDR5 memory, and other improvements over Milan-X, the HBv4 series can offer radically better performance than even the HBv3 series I was mesmerized by simply one year ago.

With the HBv4 series there can be up to 176-core VMs paired with 688GB of RAM thanks to the 96 cores per socket available with Genoa(X). All of the HBv4 VMs are currently powered by the AMD EPYC 9V33X Genoa-X processors, a custom SKU for Microsoft Azure that features a 2.4GHz base clock and 3.7GHz single-core peak frequency. Aside from the impressive CPU capabilities, the HBv4 series offers 400 Gb/s RDMA, up to 8 Ethernet vNICs, and up to 32 storage disks.

The Azure HBv4 series reached general availability state in mid-June. For getting an idea of the impact of the Azure HBv4 series, Microsoft kindly provided gratis access to their top-end Standard_HB176rs_v4 instance with 176 vCPUs. They also provided access to prior-generation HB series VMs, each at the top of stack. This comes down to testing the following configurations:

HBv4 - The 176 vCPUs with the EPYC 9V33X Genoa-X.
HBv3 - The 120 vCPUs for Milan-X with the EPYC 7V3X processors.
HBv2 - The 120 vCPUs with the EPYC 7V12 Rome processors.
HC - The original 44 vCPUs with the Xeon Platinum 8168 Skylake processors.

All of these top-of-stack HPC VMs were freshly benchmarked using AlmaLinux 8.8 with the GCC 13.1 compiler. Performance-per-dollar benchmarks are also available based on the Azure pay-as-you-go pricing, per hour for each of these instances.