AMD yesterday released a Zen 3 architecture EPYC processor code-named Milan-X with 3D vertical cache (3D V-Cache) technology. This technology can bring an additional 64MB 7nm SRAM cache for each CCD. Microsoft also uses AMD’s new EPYC processor to build the Azure HBv3 series VM. The system uses two EPYC 7773X processors with a total of 128 cores, and released relevant benchmark test results.
Among the 128 cores of this server, 8 cores are used for Azure hypervisors and other orchestration routines. Microsoft provides customers with five different core configurations, namely 16, 32, 64, 96, and 120. The highest frequency is 3.5 GHz, the allocation of the L3 cache depends on the configuration. Microsoft can upgrade the original Zen 3 architecture EPYC processor to a version with a 3D vertical cache, this process does not need to change the other configuration of the platform, including 4458GB memory and 350 GB/s bandwidth, but Microsoft still adds hardware for this, such as Mellanox ConnectX-6 NIC and two 900GB NVMe SSDs to achieve high-speed Ethernet connection (200 Gbps) and memory read/write speeds of 2.9~6.9 Gbps.
Microsoft pointed out that workloads such as CFD, FEA, weather simulation, and EDA RTL simulation all benefit from a larger L3 cache, which reduces the memory latency of the original system by 42% to 50% after the upgrade. A larger cache allows a higher cache hit rate, the combination of L3 cache and memory further improves the efficiency of the platform. Microsoft believes that in the worst case, Milan-X will be slightly slower than Milan in terms of L3 latency. According to the benchmark data provided by Microsoft, we can understand the performance difference between Milan-X compared to Milan, Rome, and Skylake. For example, in the 64-core VM, in the f1_racecar_140 benchmark model of Ansys Fluent 2021 R1, the performance of Milan-X is 77% higher than Milan and 257% faster than Skylake.