At SC23, we noticed the brand new Supermicro 4U Common GPU system. It is a liquid-cooled system designed for the densest deployments. Since we now have been doing so much with liquid cooling, we figured we’d present this off because the group goes by way of photographs from the present.
Supermicro 4U Common GPU System for Liquid Cooled NVIDIA HGX H100 and HGX 200
At SC23, we took a take a look at the brand new Supermicro 4U Common GPU system. Supermicro has quite a lot of 8U fashions which might be optimized for both air or liquid cooling, however this design is particularly designed to make the most of liquid cooling to dramatically improve density.
The liquid cooling manifold is a horizontal resolution, as we confirmed beforehand on STH. That permits the system’s cooling nozzles to be rapidly disconnected from the manifold.
On this system, the highest tray is the NVIDIA HGX H100 8-GPU with NVSwitch tray. Sooner or later, Supermicro says it’ll help the HGX H200 GPUs.
We couldn’t transfer the rack, however behind the unit, there are 4 energy provides (two put in) and an enormous set of full-height and low-profile I/O enlargement card slots. We additionally get the BMC’s out-of-band administration port, two USB 3 ports, and a VGA port.
Pulling the CPU tray out, we see a twin Intel Xeon server that’s for both Sapphire Rapids (4th Gen Intel Xeon Scalable) or the upcoming fifth Gen Intel Xeon Scalable (Emerald Rapids) CPUs. Every has a full set of 16 DDR5 DIMM slots for 32 whole.
The CPUs within the system are liquid-cooled since Intel’s socket is designed for as much as 385W TDP and sometimes higher-end CPUs are utilized in these GPU servers.
One thing our readers will discover is that there are followers on this chassis. The followers enable Supermicro to chill the DIMMs, M.2 SSDs, 2.5″ SSDs, and the rear I/O playing cards without having chilly plates on all of these units.
One can see the 2 cages for 8x 2.5″ NVMe SSDs whole on the entrance of the system right here.
Last Phrases
Total, this new system follows Supermicro’s design philosophy for AI servers, apart from the actual fact it’s primarily liquid-cooled. A 4U GPU server presents a problem for cooling as they’ll use ~10kW of energy every. Ten of those in a 45U rack could be 100kW. Utilizing liquid cooling often removes 10-15% of the facility requirement, however that’s nonetheless 80-90kW within the rack earlier than including switches.
Supermicro has a number of large-scale GPU clients that may use liquid cooling, are constructing extra energy, and wish extra density. That’s the kind of buyer this technique is constructed for.
If you wish to be taught extra about Supermicro liquid cooling, we beforehand seemed on the 8U Liquid Cooled Supermicro SYS-821GE-TNHR 8x NVIDIA H100 AI server and Supermicro’s customized liquid cooling rack.
