HPE is joining the 400Gbps generation for its Ethernet-based HPC interconnect. The new HPE Slingshot 400 not only increases the switch ASIC speed to 51.2Tbps, but it also will usher in significantly faster HPC interconnect speeds to keep pace with solutions like the NVIDIA Spectrum-X we showed in the 100K GPU xAI Colossus Cluster that are being used in AI clusters. We saw the new liquid-cooled switch on the SC24 show floor.
HPE Slingshot 400 Brings a Liquid Cooled 51.2T Switch and 400Gbps Networking
The switch itself is made for HPE’s liquid-cooled supercomputing platforms. As such, and as something different from other 51.2T switches, it is completely liquid cooled designed to not use fans. That is a challenge since it has to cool not just the switch ASIC (Rosetta 2?) but also the optics.
The back of the switch interfaces with HPE’s HPC chassis. This is very similar to how blade chassis work if you are familiar with those switches.
On the front or rear, HPE has 24 optical cages. Pluggable optics at this speed generate enough heat that these need to be liquid-cooled in a switch like this as well.
Overall, this is a very neat design and should be a big update. We also learned that there is the HPE Slingshot rebrand happening. Slingshot 10 looks to be Slingshot 100 for 100Gbps and Slingshot 11 to Slingshot 200. This is now Slingshot 400 to make things easier for folks to understand the speeds.
Final Words
It is great to see Slingshot move to the 51.2T and 400Gbps generation. At the same time, we know that Broadcom’s 102.4T switches like the Delta DC-90640 are coming out in 2025, so HPE is going to need to hustle to outpace the commodity Ethernet networking, especially as the UltraEthernet Consortium products start coming out and getting deployed into large scale AI clusters.
We heard there is a NIC somewhere and we will see if we can grab a photo or two of that later during the show.
When moving from 100 to 400 Gbps does latency also increase?