If you read our recent MI100 leak, you could probably tell that we were a little sad we couldn’t share the slides that formed the basis of the leak, mostly because many commentators questioned our leak based on the FP32 numbers. However, we have finally gotten clearance from our source(s) and determined that it’s safe to show most of the slides we received in full detail. There won’t be too much analysis here, we just wanted to show the slides themselves for the sake of brevity.
In this first slide, AMD is benchmarking a dual socket system, with each CPU having a cluster of 4 GPUs for 8 GPUs total. This is not, however, the configuration with 8 GPUs in a single cluster, that one is different. Note at the bottom that, while this system could work with Xeon, it requires PCIe switches to add more lanes. This system is expected to launch this upcoming March.
In this second slide, we can see a system that is again dual socket but only uses 2 GPU per CPU. There’s also a benchmark on the side showing that the MI100 at 300 watts is slightly faster than the A100 at 300 watts. AMD is also touting far better performance per dollar with the MI100. The Epyc version of this system is supposed to launch this upcoming December, with the Xeon version coming later in February.
AMD is pitching the MI100’s strengths here and why it makes sense for 4 specific markets: government labs, oil and gas, machine learning training, and academia.
Finally, our last slide shows how MI100 is pitted as a FP32 champion but beaten by the A100 solidly in everything else.
We didn’t get to show all of our slides, but these are most of what we got. Hopefully this adds some additional context and information for those wondering about our leak. We still stand by everything we said in our original article, so please read that if you want the full story.