According to Phoronix, Microsoft’s Azure HBv5 virtual machines powered by custom AMD EPYC 9V64H processors with HBM3 memory have finally reached general availability nearly a year after their initial announcement. These VMs feature up to 368 AMD Zen 4 cores clocking up to 4.00GHz and deliver a staggering 6.9TB/s of memory bandwidth across 400-450GB of HBM3. The flagship configuration provides up to 9GB of HBM3 memory per core and runs without SMT enabled. Microsoft provided Phoronix with early access to benchmark these new instances against the previous HBv4 series powered by AMD EPYC 9V33X Genoa-X processors with 3D V-Cache. Both systems were tested running Ubuntu 24.04 LTS with Linux kernel 6.14 across a wide variety of HPC workloads. The results show the EPYC 9V64H delivering impressive performance for memory-intensive applications despite being based on Zen 4 architecture rather than the newer Zen 5.
Sponsored content — provided for informational and promotional purposes.
The Timing Trade-Offs
Here’s the thing about this launch – it’s both impressive and slightly behind the curve. We’re talking about Zen 4 architecture when Zen 5 has been available for over a year already. That’s an eternity in the chip world. But honestly? The HBM3 memory bandwidth of nearly 7TB/s is so massive that it probably makes the architecture generation less critical for the specific workloads these VMs target.
I can’t help but wonder why Microsoft didn’t wait for Zen 5 + HBM3. The design and validation cycles for these custom server chips must be brutal. Maybe they had to ship what was ready rather than perfect. Or perhaps the performance uplift from HBM3 alone was so significant that waiting another year for Zen 5 didn’t make business sense.
Who Actually Needs This?
Let’s be real – these aren’t your everyday cloud instances. We’re talking about specialized HPC workloads where memory bandwidth is the bottleneck. Think computational fluid dynamics, financial modeling, weather simulation, that sort of thing. For regular web servers or databases? Complete overkill.
The fact that Microsoft is offering up to 9GB of HBM3 per core tells you everything about the target audience. These are researchers and enterprises running simulations that would choke standard cloud instances. And at 368 cores per server without SMT, they’re clearly prioritizing raw single-thread performance over core count.
The Cloud HPC Arms Race
This launch isn’t happening in a vacuum. Every major cloud provider is pushing deeper into specialized HPC offerings. AWS has their Graviton chips, Google has TPUs, and now Microsoft is doubling down on AMD partnerships with custom silicon. It’s becoming less about generic compute and more about workload-specific optimization.
What’s interesting is that we’re seeing cloud providers move beyond just slapping standard server CPUs into virtual machines. They’re now co-designing processors specifically for their infrastructure and customer needs. That’s a significant shift in the cloud computing landscape. Basically, the cloud wars are moving from who has the most data centers to who has the smartest silicon.
A Glimpse of What’s Next
Looking at these benchmarks, the performance uplift from HBM3 is undeniable for memory-bound workloads. But this feels like just the beginning. Imagine what happens when Zen 5 architecture gets paired with this memory technology. Or when other cloud providers respond with their own HBM-powered offerings.
The real question is how quickly this trickles down to more mainstream cloud instances. Probably not anytime soon given the cost, but the performance characteristics we’re seeing here will influence future designs across the board. For now, if you’re running memory-intensive HPC workloads, these Azure HBv5 instances look like they’re worth the wait.
