The AMD Instinct MI200 Series Accelerator

The marketing material is cool and all, but article form is just easier to discuss and quote.

https://www.anandtech.com/show/17054/amd-announces-instinct-...

TL;DR: MI200 presents itself as 2x GPUs to the OS. Together, they push 47.9 TFLOPS double-precision AND single-precision (yes, FP32 does not seem to benefit. So AMD is focusing a huge amount of effort on maximizing FP64 performance)

---------

Anandtech points this out in a cheeky paragraph:

> This gives each GCD about 1.64TB/second worth of memory bandwidth, or as AMD likes to promote it, a cumulative memory bandwidth of 3.2TB/second.

So 3.2TB spec isn't wrong per se, but its definitely modern marketing at work. It doesn't look reasonable for a single-kernel to ever get more than 1.64 TB/s, but GPUs are parallel machines, and we can reasonably expect multiple kernels to be executing across the two GCDs in parallel.

But this MI2xx series will function more like two tightly linked GPUs, rather than a singular GPU-instance. Fortunately, most GPUs deployed today on supercomputers are 6-GPU or 8-GPU clusters.

So this "two GPUs glued together" package is just what the supercomputer world needs. This was tried years ago for video-gamers, but the latency just was too much. But supercomputer workloads have ways around latency (mostly by just... not caring about it because it takes 2+ weeks for a lot of calculations to run)