AMD MI300As power French Adastra supercomputer!

0

We’re heading for France, and more precisely to CINES in Montpellier, where we’ll find a supercomputer: Adastra. It’s powered by AMD’s latest Exascale-class APUs: the MI300A. It ranks 17th among the world’s most powerful supercomputers, with a computing power of 46.1 PFLOP!

MI300A powering Adastra!

Adastra supercalculateur

The machine configuration features HPE Crat EX4000 cabinets. These contain a total of 14 HPE Cray EX255a blades featuring the famous MI300A APUs. Finally, to interconnect all this, we find HPE Slingshot switches enabling 200 Gbps connections per port. All in all, the first tests in November revealed a total computing power of 46.1 PFLOP. Just to clarify, this supercomputer will be used for research into climatic environments, astrophysics, chemistry and so on.

What’s a MI300A?

Simply put, a MI300A APU is an APU combining a very large graphics unit with CPU cores. In this case, the configuration mixes Zen4 cores via CCDs similar to those of the Epyc 9004 and a massive CDNA3 GPU. We’re talking about an iGPU with 228 Compute Units for a total of 14,592 Stream Processors. The CPU is made up of three CCDs, each containing eight Zen4 cores for a total of 24.

AMD Instinct MI300A

Finally, let’s talk about the memory, which is also colossal. We’re talking about 192 GB of HBM3 memory, coupled with an 8192-bit bus. The result is a bandwidth of 5.3 TB/s.

As for frequencies, we find a GPU running at 2100 MHz peak, while TDP peaks at 760W (base at 550W). AMD has announced performance figures of up to :

  • Single Precision Matrix (FP32): 122.6 TFLOPs
  • Double Precision Matrix (FP64): 122.6 TFLOPs
  • Half Precision (FP16): 980.6 TFLOPs

It goes without saying that this model benefits from a chiplet design that combines several etchings: 5 nm and 6 nm. Finally, this model contains 153 billion transistors.

It should also prove popular, with sales estimated at between 300,000 and 400,000 units by 2024. What’s more, Microsoft Azure will be the first cloud service to run on this hardware.

Clearly, between Intel and AMD, NVIDIA has its work cut out, and may no longer have a monopoly… On the other hand, when you see the waiting times, you’d think there’d be plenty of room for other players.