HPE Launches New AMD EPYC Venice Instinct MI400 and NVIDIA Vera Rubin Compute Blades

As we prepare for Supercomputing 2025 next week, HPE has launched a trio of new blades for its GX5000 compute platform. Two of the compute blades feature next-generation AMD EPYC “Venice” CPUs, with one of these two blades also featuring AMD Instinct MI400 GPUs. The third blade is a next-generation NVIDIA Vera Rubin design.

HPE Launches New AMD EPYC Venice Instinct MI400 and NVIDIA Vera Rubin Compute Blades

Here are the details on the new blades:

• HPE Cray Supercomputing GX440n Accelerated Blade – This blade features four NVIDIA Vera CPUs and eight NVIDIA Rubin GPUs. HPE says they can fit 192 NVIDIA Rubin GPUs per rack, with up to 24 of these blades per rack.
• HPE Cray Supercomputing GX350a Accelerated Blade – This blade features a single Zen 6-based AMD EPYC Venice CPU and four AMD Instinct MI430X GPUs. HPE says they can fit 28 of these per rack, which would be 112 AMD Instinct MI430X GPUs. Notably, the AMD Helios liquid-cooled racks shown to date also have one EPYC Venice with four MI450X accelerators. The MI430X is tailored for HPC, while the MI450X is focused on AI.
• HPE Cray Supercomputing GX250 Compute Blade – This blade features only the Zen 6-based AMD EPYC Venice GPUs. HPE says that they can fit 40 of these compute blades in a single rack.

Notably absent from this list is an Intel Xeon blade, after Intel won one of the first Exascale computing systems with Aurora. While Aurora also used Intel’s GPUs, it is notable that HPE made a multi-vendor compute blade announcement and left out Intel’s Diamond Rapids generation Xeons from a CPU-only solution. Perhaps that is just another sign of the times with CPU-GPU co-design happening.

We covered the new HPE Cray GX5000 series launch as part of our ORNL Discovery and Lux Powered by HPE and AMD Announcement recently. One of the new innovations is that on the liquid cooling side, HPE can support heterogeneous blades. A reason to do that might be to support CPU-only blades along with GPU-accelerated blades. Now that we have some more information on some of the blade options, perhaps that makes sense.

HPE also said that its 64-port 400Gbps Slingshot blades will be completely liquid-cooled. HPE is offering 8, 16, and 32 switch configurations, which means 512, 1024, and 2048 ports. What is interesting is that since the original Slingshot launched, the AI build-out has pushed Ethernet technology to the forefront, including Broadcom Tomahawk Ultra for scale-up Ethernet clusters and enormous multi-site AI clusters. It will be interesting to see how the AI industry’s use of Ethernet across scale-up, scale-out, and scale-across domains impacts HPE’s focus on Slingshot going forward.

HPE has been doing liquid-cooled Slingshot switches for some time now. Above is a photo we took from El Capitan’s Rabbit and Slingshot networking stack.

Final Words

Our hope is that we find some of these blades at SC25 this year. HPE said that the new compute platforms and Slingshot 400 are scheduled for early 2027. That aligns with the timelines set forth in forums like the AMD Financial Analyst Day 2027 earlier this week for next-generation chips arriving in 2026. Given the scale and speed of the AI build-outs, we would expect next-generation chips to arrive for AI clusters before HPC.

If you want to see the current-generation HPE Cray El Capitan, we were able to tour it just before it went classified earlier this year.