Microsoft convinced AMD and Nvidia to build a CPU with extraordinary features, but it will never go on sale: the 4th generation 9V64H has 88 cores and uses InfiniBand technology
- Microsoft has unveiled a specially designed AMD CPU
- It is only available on Azure and uses ultra-fast HBM memory
- Nvidia’s InfiniBand technology has also been used
Earlier today at Ignite, Microsoft CEO Satya Nadella pulled back the curtain on a new AMD processor that’s destined to become the most powerful CPU to run on Azure cloud computing infrastructure, one that comes with a few surprises. You can request sample access for the Azure HBv5 VMs by completing this form.
The 9V64H is a custom 4th generation EPYC server CPU, not one based on the latest 5th generation, codenamed Turin. It has 88 cores and is clocked at up to 4GHz; the closest existing model is the 9634, an 84-core CPU with a top speed of 3.7 GHz.
According to Fernando AznarHPC + AI Product Marketing Manager at Microsoft, it will focus on “most memory bandwidth-intensive HPC applications, such as computational fluid dynamics, automotive and aerospace simulation, weather modeling, energy research, molecular dynamics, and computer-aided engineering.” But not a word about AI.
What makes the CPU What’s so special is that a cluster of four of these, as part of Azure HBv5 VMs (launching in 2025), will have access to up to 450GB of RAM, in that case HBM3, with a combined memory bandwidth of almost 7.0TBps.
That’s more than the AMD Instinct MI300A accelerator without the GPU part (24 cores, 3.7 GHz peak, 128 GB HBM memory, 5.3 TBps peak performance) and almost an order of magnitude higher than the rest of the competition (including AMD’s).
A very unique CPU
But there’s more to the story: this is AMD’s first quad-socket project, which is an interesting take on what AMD told me earlier this year when I interviewed Robert Hormuththe company’s corporate vice president, architecture and strategy.
Clearly the market is looking for single socket, unless there is a clear reason – and a hyperscale customer with virtually unlimited resources – to go for quads.
Microsoft also confirmed that the chip is SMT-enabled (just like the Zen 4c and Zen 5c parts) and that the VM will have access to as much as 3.17 TB of memory (up to 9 GB per core).
And while this is a co-engineered chip, Microsoft is firmly in the driver’s seat by enlisting Nvidia, AMD’s nemesis, to provide industry-leading Quantum-2 Infiniband networking technology to power CPUs with bits (800 Gb/s , split evenly across four CPUs).
This, Microsoft says, will give customers the ability to “efficiently scale their workloads to hundreds of thousands of cores” as needed.
I was also intrigued by the mention of a 14TB local NVMe SSD that delivers up to 50GB/s read and 30GB/s write performance, which is far better than anything using PCIe Gen5 technology and can only be achieved using specialized hardware such as Highpoint’s Rocket 1608A expansion card.
