This week Supermicro announced a number of new NVIDIA MGX systems. We knew the release was inbound, but then we managed to find one at OCP Summit 2023. That was an especially interesting model with the upcoming NVIDIA GH200 Grace Hopper part. As one would expect, we have photos.
Supermicro ARS-111GL-NHR 1U Grace Hopper Server
Here is the overview card for the server. We can see that not only do we get the 72 core Arm Neoverse V2 processor with 480GB of LPDDR5X. There is also an integrated H100 with 96GB HBM3, more than the standalone version’s current 80GB total. In addition, we get features like 3x PCIe Gen5 x16 slots and a variety of storage options.
On the front, there are eight E1.S slots.
Then we immediately transition to cooling for the GH200 with a giant heatsink.
Here is a look at the heatsink that extends well past the GH200 module and motherboard in order to cool NVIDIA’s part in a 1U chassis.
On the side of the GH200, we get a vertical M.2 riser board.
At the rear of the system, we get the full-height expansion cards. This system was not limited to just the 72 Arm cores on the GH200. Instead, there are two other Arm-based systems running on the BlueField-3 DPUs at the rear of the system.
A few things to notice in this chassis is that the motherboard is relatively compact as it is mostly cabled connections to all of the other components. The other part is that with the NVIDIA MGX platforms we have seen to date, there tends to be a ton of room dedicated to airflow instead of maximizing other components in the server. It takes a lot to cool such a hot spot like a GH200 in a 1U server and then still cool power-hungry BlueField-3 cards so that is why we have so much room for cooling in MGX platforms like this.
Beyond this ARS-111GL-NHR, Supermicro also launched a number of other NVIDIA MGX platforms. Here is the full list:
- ARS-111GL-NHR – 1 NVIDIA GH200 Grace Hopper Superchip, Air-Cooled
- ARS-111GL-NHR-LCC – 1 NVIDIA GH200 Grace Hopper Superchip, Liquid-Cooled
- ARS-111GL-DHNR-LCC – 2 NVIDIA GH200 Grace Hopper Superchips, 2 Nodes, Liquid-Cooled
- ARS-121L-DNR – 2 NVIDIA Grace CPU Superchips in each of 2 Nodes, 288 Cores in total
- ARS-221GL-NR – 1 NVIDIA Grace CPU Superchip in 2U
- SYS-221GE-NR – Dual-socket 4th Gen Intel Xeon Scalable processors with up to 4 NVIDIA H100 Tensor Core or 4 NVIDIA PCIe GPUs
Something one will quickly notice here is just how many liquid-cooled Grace Hopper servers there are on the list. Supermicro has been working on its own custom liquid cooling platform that you can learn more about in the video.
We already covered the manifold and CDU part of this solution and will have a few more web-only pieces in the near future. Hopefully, this MGX list explains why we are putting such an emphasis on liquid cooling at STH.
Final Words
The NVIDIA Grace Superchip and Grace Hopper parts are cool, but the question is whether they will sell. We know that Supermicro was already seeing interest in these platforms before they launched. In an environment of huge demand for NVIDIA’s products some are striking out for the Grace Superchip and Grace Hopper as a way to hopefully get next-generations of compute sooner.
