2 ExaFLOPs Supercomputer: El Capitan is equipped with the Instinct MI300 APU with AMD socket

Image: AMD

It was known that El Capitan would use Zen 4 and a new generation AMD Instinct. It is now certain: it will be the HP-APU Instinct MI300 based on Zen 4 and CDNA 3. The supercomputer, which will also simulate nuclear weapons, will be built from next year and will be launched in the first half of 2024.

More performance also thanks to more power?

The exact specifications for the system are not yet available, but two years after the first information, some parameters have been announced or confirmed again. The following also applies: More than two 2 ExaFLOPS FP64 performances should be achieved, which is almost double (or maybe later) than the current Frontier design.

In March 2020, however, there was only discussion of a maximum of 30 megawatts, but today makers are asking for up to 40 megawatts, which should not be broken. Frontier got 1.1 exaFLOP at 21.1 megawatts. On this point, the manufacturers of the next supercomputer risk being overwhelmed before the final hardware is available, because otherwise the increase in performance and consumption would be almost linear. However, this can hardly be the goal of a next-gen solution and should therefore probably be surpassed eventually in terms of efficiency.

With “HPC-APU” from AMD instead of CPU and GPU in a ratio of 1: 4

Two years ago, AMD explained that four GPUs and one Epyc CPU would be paired per node. Additionally, AMD said at the time that the third generation of the Infinity architecture would be used, with completion scheduled for early 2023. But planning is history, and so is the hardware, at least in part.

The delay in delivery and use since Q2 2024 appears to allow for even newer hardware and previous information is out of date.

Because the AMD Instinct MI300 is already based on the fourth generation Infinity architecture, not the previously advertised third generation, including stacking, that is, not only connected in the layer, but also chips stacked on a single package.

AMD promises 8x higher AI performance than Instinct MI200
AMD promises 8x faster AI performance than Instinct MI200 (Image: AMD)

AMD has officially announced a socketed APU on this base under the Instinct MI300 brand, which consists of a Zen 4 CPU and a CDNA 3 GPU, which can access shared HBM3 memory (Unified Memory). CPU and GPU are placed on an infinite cache on the package. It is not yet known how many cores the CPU has or how many GPU cards the GPU has. The fact that the APU can be plugged into the socket facilitates integration into nodes.

The editors of ComputerBase would like to thank the ETI1120 community member for reporting this message.

About the author


Leave a Comment