Sizing up MI300A’s GPU - by Chester Lam - Chips and Cheese
AMD’s Instinct MI300A is a giant APU, created by swapping out two GPU chiplets (XCDs) for three CPU chiplets (CCDs). Even though MI300A integrates Threadripper-like CPU muscle, the chip’s main attraction is still its massive GPU compute power. Here, we’re going to size up MI300A’s GPU and see where it stands using comparison data from a few other GPUs, including MI300X.
Special thanks goes out to AMD and GIGABYTE along with their Launchpad service , who generously let Chips and Cheese play around with a massive quad-socket MI300A system system in the form of the G383-R80-AAP1 for over 2 weeks. As always, our testing was our own.
We also have limited data from Nvidia’s H100. H100 comes in both a PCIe and SXM5 version. I (Clam) rented a H100 PCIe cloud instance for the MI300X article from before. Cheese/Neggles rented a H100 SXM5 instance for benchmarking against the MI300A.
MI300A may be down a couple XCDs compared to its pure GPU cousin, but it still has plenty of compute throughput. It’s well ahead of Nvidia’s H100 PCIe for just about every major category of 32-bit or 64-bit operations. H100’s SXM5 variant should slightly increase compute throughput, thanks to its higher SM count. But a 16% increase in SM count likely won’t do much to close the gap between H100 and either MI300 variant.
