Supermicro showed off a large number of systems at ISC 2023. One of them particularly will be impactful. Supermicro had a liquid-cooled 8x NVIDIA H100 Delta-Next platform at the show and it was different than we expected.
Supermicro Shows Liquid-Cooled NVIDIA H100 Delta Next at ISC 2023
First off, we are just going to point out that Supermicro was showing off a huge array of accelerated systems. We may show those other systems later, but something that Supermicro was showing beyond many other vendors like Dell, HPE, Lenovo, and more was the UBB acceleration platform. UBB is a child of the OCP project and is the accelerator platform that accepts OAM accelerators like the AMD MI250X, Intel Habana Gaudi 2, Intel Data Center GPU Max (Ponte Vecchio) and more.
Supermicro also had the NVIDIA H100 air-cooled Delta Next platform. We recently covered the Delta Next HGX H100 platform. We covered the NVIDIA DGX versus NVIDIA HGX What is the Difference previously.
The Supermicro SYS-821GE-TNHR-LCC was the system at the show. Here we can see the system which shares a chassis with the air cooled version. If you saw the X13 GPU accelerated systems launch video, you may have seen this platform.
The top of the system has the GPU tray. The bottom section has the SSDs as well as the fans, CPUs, and NICs.
Opening up the top tray of the chassis, we have a large shelf with the GPUs and the liquid cooling blocks.
The liquid cooling blocks cover not just the NVIDIA H100 GPUs but also the NVLink NVSwitches.
Lighting was rough at the show, but here is another view of these where you can see the GPUs on the left and middle, the NVSwitch on the middle right, followed by the liquid tubes on the far right.
The blocks were branded by Supermicro. This was a bit different since the NVIDIA HGX H100 / Delta Next platform has heatsinks that NVIDIA sells already on the GPUs. Supermicro is making these liquid cooling blocks.
Supermicro also has a quick disconnect manifold and fittings. We saw them disconnected to pull out the GPU tray, and they were very fast to pop on and off.
Below the GPU systems in the liquid cooled rack, there was also a Supermicro designed CDU with hot swap pumps.
We showed the previous gen NVIDIA A100 liquid cooling solutions and this is certainly a few generations newer.
The key learning with this is that not only is Supermicro selling air-cooled NVIDIA H100 systems, but this is the next step of engineering. With the accelerated systems, there is a major push towards liquid cooling. The ISC 2023 announcement said that Supermicro has the first liquid-cooled HGX H100 system so this is just an example of where the company is a bit ahead on the AI front.
Now, the big challenge is really H100 availability.
Their rack nicely exhibited one common issue that I’m facing myself: most of these new power hungry gpu machines come with 16A c19/c20 power plugs while common rack PDUs are predominantly 10A c13/c14 based. Looks like that before we start to deploy such machines, we have to at least rethink our PDUs and power delivery.
Another option uses ambient air and no liquids. Forcedphysics.com