At SC23, we saw a number of next-gen AMD systems. While we know HPE-Cray and AMD are installing El Capitan a 2 Exaflop Supercomputer, we also got news this week that Microsoft was deploying large numbers of AMD’s MI300X GPUs. We thought folks may want to see the new high-end GPUs and APUs, so we snapped a few photos.
AMD MI300 Recap
For those who need a quick refresher, the AMD Instinct MI300 family is a modular building block architecture where AMD combines GPU, CPU, HBM, and more into a 3D stacked package. AMD can vary the GPU and CPU tiles to achieve different types of configurations. Since AI is selling like hotcakes right now, the two most popular configurations we are likely to hear about in the near future are the MI300A which is one-quarter CPU and three-quarters GPU, and the MI300X which is all GPU.
At SC23, we saw both.
AMD Instinct MI300A at SC23
First is the AMD Instinct MI300A. This is a large APU that we grabbed a photo of in June, but for some reason the dimensions did not register.
We saw the APU on the show floor in the Gigabyte G383-R80. One can see that this is a socketed processor.
The socket looks very familiar even though it is labeled “SH5”.
Indeed, these even use the same style of carrier that we see in AMD’s mainstream server CPUs today.
The carrier is grey much like the AMD EPYC Milan-X.
Our Genoa-X samples had orange carriers like Genoa (green is Bergamo.) Perhaps grey is AMD’s color for its high-end HPC CPUs?
These were installed into the LGA6096 sockets in a Gigabyte G383-R80. One may recall that socket SP5 for the AMD EPYC Genoa, Genoa-X, and Bergamo is a LGA6096 design as well. Four of these sockets were on the motherboard.
What we are most interested in getting more details on is the CXL support. These chips do not have DIMM slots and are instead using onboard HBM for memory needs, much like the Intel Xeon MAX 9480, except with more capacity and as mostly accelerator. The chips have multiple PCIe Gen5 interfaces. Before AMD’s June 2023 data center event, we heard folks interested in memory expansion via Type-3 CXL devices to get more memory capacity. AMD has still not said whether this is a supported feature.
AMD also had the MI300X at SC23, so let us get to that next.