Microsoft Azure Unveils HBv5 VMs with Custom AMD CPUs for High-Performance Computing

  • Thread Author
In a move that's likely to send ripples through the high-performance computing (HPC) landscape, Microsoft has unveiled its latest Azure HBv5 virtual machines, powered by a cutting-edge custom AMD CPU. This new CPU, which sports an impressive configuration, seems to be a rebirth of the long-rumored MI300C design, once relegated to the speculation chamber of tech enthusiasts.

Unpacking the CPU Specifications​

So, what precisely are we looking at with this new custom AMD EPYC CPU? Brace yourselves. Microsoft has engineered a chip that features 88 Zen 4 cores and boasts 450GB of HBM3 memory. When leveraged across four chips, these VMs unleash a staggering bandwidth of 7 TB/s, which Turbocharges performance compared to their predecessors.
Previously utilitarian CPUs like the Milan-X and Genoa-X provided some impressive bandwidth via AMD's 3D V-Cache technology. Still, with the HBv5 series, Microsoft has decided to crank things up a notch. They recognized that memory bandwidth was a critical bottleneck in HPC workloads—a detail that many may overlook while contemplating the intricacies of CPU architecture.

What's the Big Deal About HBM3?​

HBM3 (High Bandwidth Memory) is the spry new entrant in memory technology. Typically reserved for data center-grade GPUs, it elevates the game by offering higher bandwidth and lower latency compared to traditional DRAM. In the context of Azure's HBv5 VMs, HBM3 serves a dual purpose: it provides substantial memory bandwidth while allowing the CPUs to act more efficiently by relocating communication needs away from conventional RAM.

Performance Metrics Ramifications​

Let’s talk performance. Bolstering these Azure VMs with HBM3 gives them a remarkable bandwidth increase of nearly nine times over the previous Genoa-X chips and an almost 20-fold improvement over the Milan-X counterparts. The implications for users who rely on Azure for compute-intensive tasks—such as machine learning and large-scale simulations—are monumental.
The HBv5 VMs are equipped with 352 Zen 4 cores (though Microsoft has wisely disabled SMT for this configuration), ensuring that even the most demanding workloads can run with efficiency and speed. Not only will organizations be able to perform computations faster, but Azure's offerings will likely become more appealing as businesses scramble to leverage cloud computing for their computational needs.

A Custom CPU That May Not Be So Custom​

Intriguingly, this so-called custom CPU isn’t as bespoke as one might think. It reflects characteristics akin to AMD's MI300C—a chip designed but never officially released. The MI300C was initially conceptualized with the intent of merging CPU and GPU technology, featuring Zen 4 cores instead of CDNA architecture. This blurring of lines between traditional computing roles aligns well with current trends in processing technology, where workloads often demand hybrid capabilities.
Phil Park, an AMD memory engineer, shed light on the challenges faced by high-bandwidth memory designs. According to him, the commitment to HBM3 means AMD has had to make specific architectural decisions, emphasizing that HBM does not suit all CPU applications, given its bandwidth requirements. This context casts a reflective light on why market deployment has been slow in adopting HBM-equipped CPUs: they represent a considerable design and manufacturing investment, rather than a flexible mainstream offering.

Azure's Exclusive Offering: A Strategic Move​

What makes this CPU particularly noteworthy is its exclusivity to Microsoft Azure. Glenn Lockwood from Microsoft confirmed that this chip would not be available for any other distributors or as standard EPYC CPUs. This exclusivity indicates a significant strategic play by Microsoft—possibly a means to tilt Azure’s competitive balance in cloud services. By collaborating closely with AMD to develop this solution, Microsoft is not merely aiming to keep pace with competitors like AWS and Google Cloud; they’re looking to set the standard for what high-performance computing should look like.

Conclusion: Bridging the Future of HPC in the Cloud​

This latest iteration of Microsoft's Azure offerings solidifies its place in the expanding cloud landscape of HPC. By harnessing the might of AMD's HBM3-enabled custom EPYC CPUs, users can expect phenomenal advancements in computational capabilities that were previously confined to academic and research institutions.
As we anticipate further developments, it will be exciting to see how businesses will adapt to and implement the new capabilities afforded by Azure's HBv5 virtual machines. The era of ultra-efficient, memory-bountiful cloud computing is upon us, and the future looks incredibly bright for Windows users willing to tap into Azure's newest offerings. So, as you contemplate your next computing needs, remember: the frontier of high-performance computing is evolving, and it’s happening right in the cloud.

Source: Tom's Hardware AMD crafts custom EPYC CPU for Microsoft Azure with HBM3 memory – CPU with 88 Zen 4 cores and 450GB of HBM3 may be repurposed MI300C, four chips hit 7 TB/s
 


Back
Top