AMD creates custom EPYC CPU with HBM3 memory for Microsoft Azure
AMD crafts custom EPYC CPUs for Microsoft with Zen 4 cores and HBM3 memory on-chip – Insane bandwidth
Microsoft has revealed their latest HBv series virtual machines for Azure, promising incredible performance levels thanks to their custom AMD EPYC CPUs. These customer EPYC CPUs feature an exclusive design with HBM3 memory, up to 352 Zen 4 CPU cores, 6.9 TB/s of memory bandwidth.
Microsoft’s new HBv5 virtual machines can feature between 400 and 450GB of HBM3 memory. These VMs have four processors total, each with 88 Zen 4 CPU cores. That gives this CPU up to 9GB of HBM3 memory per CPU core, which is a lot of memory to play with. Additionally, this memory provides users with faster access than standard DRAM, as this HBM memory is connected directly to the CPU through an interposer.
EPYC CPUs with HBM3 memory
The key feature of Microsoft’s new HBv5 VMs is their memory bandwidth. Memory performance can be a huge limiting factor for enterprise CPU users. That’s why AMD and Microsoft co-developed these CPUs to overcome these potential bottlenecks. Microsoft claims that their new Azure HBv5 systems are up to 8x faster than their competitors in terms of memory bandwidth. Additionally, they claim that these VMs are up to 35x faster than 4-5 year old HPC servers, which are entering the end of their hardware lifecycle.
These new VMs from Microsoft deliver astounding levels of memory bandwidth. Simply put, the bandwidth offered by these new VMs makes all of Microsoft’s prior HBv virtual machines look slow by comparison.
Microsoft’s new HBv5 server CPUs reportedly feature double the Infinity Cache bandwidth of standard AMD EPYC CPUs and 800Gb/s Nvidia Quantum-2 InfiniBand for network switching. These CPUs also have SMT disabled and are designed for single-tenant use to maximise security.
In a sense, AMD’s use of HBM3 memory solves a similar problem as their 3D V-Cache technology. Both solutions place fast memory closer to the CPU than traditional DRAM. While V-Cache adds memory directly to the CPU in the form of extra L3 cache, HBM4 acts as a (kind of) L4 cache with memory connected to the CPU through an interposer. HBM3 memory provides more bandwidth than standard DRAM and is much less latent. This allows CPU cores to be fed with data faster, accelerating many workloads.
You can join the discussion on AMD’s Custom EPYC CPUs for Microsoft’s Azure HBv5 VMs on the OC3D Forums.
