HPE was showing off a server blade from its upcoming El Capitan supercomputer at the recent ISC High Performance event in Hamburg, Germany.
2VIEW GALLERY – 2 IMAGES
The server blade had its front cover removed, showing off the internals that include AMD’s new Instinct MI300A AI accelerator. The blade is called the HPE Cray Supercomputing EX255a accelerator blade, featuring a single-slot 1U blade chassis. It might be small, but it packs an incredible 8 x MI300A chips, with copper cooling blocks and copper cooling pipes linking everything together, rather beautifully I might add.
Each of the blades uses liquid cooling to keep the crazy amounts of heat generated by the 8 x AMD Instinct MI300A chips, with each of the MI300A APUs featuring a 550W TDP, and a peak power rating of 760W. This means that the cooling for the blades needs to be able to handle an incredible 4400W of power on average, and peak power of up to an even higher 6080W.
Each blade features two 4-socket node cards (boards) with two MI300A APUs per card, each of the blades features an additional NVMe M.2 SSD if required, with 4-8 injection ports that are designed to connect to El Capitan’s HPE Slingshot-11 networking system.
The new HPE El Capitan supercomputer will be deployed later this year, and once it is up and running, it’ll be the world’s fastest supercomputer. It will take the new #1 spot away from the AMD-powered Frontier supercomputer.
AMD’s new Instinct MI300A APU features 24 Zen 4-based CPU cores, and a CDNA3-based GPU with 224 Compute Units and 14,592 Streaming Processors. The CPU and GPU have shared access to 8 stacks of HBM3 memory with up to 128GB of HBM3 at the ready. The new MI300A is the largest chip that AMD has ever produced, with 9 linked compute dies (1 x CPU, 8 x GPU) in total.
AMD is using TSMC’s 5nm process node for the CPUs and GPUs on the Instinct MI300A APU, while 6nm-based dies are used for the 3D-stacked dies.