At CES 2023, AMD disclosed more details on its new supercomputing product, the AMD Instinct MI300. It did a bit more than that, actually showing off the chip on stage and talking about its makeup. The AMD Instinct MI300 is designed to bring CPU cores, GPU compute, and memory all onto a single package. This eliminates enormous amounts of long-distance DDR memory trips and CPU-GPU PCIe trips thereby improving efficiency. This is the future of supercomputing. Let us get into it.
AMD Instinct MI300 Breaks Cover at CES 2023
We have discussed the AMD Instinct MI300 before, but we have details beyond the 128GB of HBM3 and 146B transistors. This includes the fact that the package will utilize 5nm and 6nm IP and that it will have 24 CPU cores.
Here is the actual package being shown by Dr. Lisa Su, AMD’s CEO. This is a large chip.
AMD says it is using 3D stacking to put multiple CPU and GPU chiplets on top of a base die that connects to memory and the rest of the system. AMD says it has nine 5nm chiplets that are 3D stacked, atop four 6nm chiplets with HBM surrounding it. The total is a 146 billion transistor part.
AMD says that the new part has much higher AI performance than the MI250X.
Final Words
AMD says it will be available in the second half of 2023 for HPC and AI solutions. That will put it likely after NVIDIA Grace, NVIDIA’s Arm CPU cores plus Hopper GPU IP, but before Intel’s Falcon Shores unified product. This new generation should be a solid step-function in performance by eliminating costly data movement in systems by co-packaging CPU, GPU, and memory. Hopefully we will get to see one sooner rather than later!
Begs the question: Is El Capitan going to snatch the supercomputing crown with MI300 before Aurora even has a chance to conquer it?
expect trickle down chiponomics from this waif to prosumers in 36-48 mos #the brady bunch #james brady
Aurora should be on the June list, shouldn’t it? El Capitan is slated for a late 2023 delivery, which means it wouldn’t be on before November 2023, and perhaps June 2024.
Besides, is it even certain that El Capitan will be faster than Aurora? Both are talking about “>2 exaflops peak performance”. Aurora’s power envelope is <=60 MW while El Capitan's is <=40 MW. I'd expect El Capitan to be faster, but I think it remains to be seen.
Will these be deployed as 1P/2P servers or more exotic configurations?
Fascinating. I wonder how this will be presented. As an add-on accelerator module, or whether it would be able to run in standalone mode, which would likely make it the fastest SoC to date. And if course, if it’s the latter, whether it can run Crysis? Though I have doubts about it having display outputs in the first place.