GPUsNews

NVIDIA CEO Brushes Off AMD’s MI200 MCM GPU: Every year there is an Nvidia killer and people call it that

AMD recently announced its Instinct MI200 “Aldebaran” data center GPU, offering nearly 5x more FP64 throughput compared to NVIDIA’s top A100 Tensor Core Accelerator. This, of course, means NVIDIA needs to react to the development in one way or another. In an interview with TNP, NVIDIA CEO Jensen Huang dismissed the Instinct MI200 as just another competitor in the data center and HPC market.

Timothy Prickett Morgan: Now there is a two-year cadence for GPUs, DPUs and soon also for CPUs that companies can count on.

Obviously, AMD is much more competitive with its Instinct MI200 “Aldebaran” series accelerators than it has ever been. It’s actually two GPUs, not one, and I’ve reminded everyone that AMD had “made a K80” by putting two GPUs in one device, but still, this GPU has won two ex-scale class systems and many more small systems.

I am aware that there will be no new GPU announcements from Nvidia until next year, based on cadence, but what is your response to this competition from AMD, and soon, to a lesser extent, Intel in the field of GPU computing?

Jensen Huang: First of all, we have competition all the time. So it is not true that this is the first “Nvidia killer” to hit the market. Every year there is an Nvidia killer and people call it that .

Timothy Prickett Morgan: I’m talking about the high-end AI and HPC supercomputer space. For the past decade and a half, when it comes to GPU-accelerated supercomputers, you’ve been the only player.

Jensen Huang: Actually, I think this is the easiest space, and let me tell you why. The reason is that an HPC machine needs two things, only two. HPC centers order everything years in advance, so they have no idea what the performance of a given device will be, but here’s the equation for you …

Timothy Prickett Morgan: Go ahead …

Jensen Huang: The number of maximum FP64 FLOPs and the memory capacity, put those two things in a cube. And in the other bucket, put dollars. That is all. That is the equation. And you know that …

Timothy Prickett Morgan: And so AMD decided to increase FLOPs and cut the price? That’s what I think happened …

Jensen Huang: The question is how we see the world, and why the competition is so intense for us. And it is very intense. It is not of less intensity. It’s seriously intense. Accelerated computing is not for people with a weak heart. So let me prove it.

You can build the best chip in the world for everything, you put it in the computer, and what are you going to accelerate? Absolutely nothing. It is not true? Accelerated computing is incredibly difficult. And the reason is that Moore’s Law is incredibly good. No one has looked at Moore’s Law, even at its slow pace, and said over time that it is not one of the most formidable technological forces in human history. And yet, for us to be successful as a company, we have to deliver results well above Moore’s law.

The Next Platform (Via: EC)

Stating that there’s competition all the time, Huang explained that the MI200 isn’t the first “NVIDIA killer” to hit the market. He went on to say that there’s an NVIDIA killer every year and people call it that. Team Green’s CEO believes that the high-end AI and HPC space is the easiest, requiring “just” high FP64 compute performance and plenty of memory and bandwidth.

This image has an empty alt attribute; its file name is FDr0bCGWEAAgA1G-1-1024x582.jpg

However, if we look at the capabilities of the MI200, it’s way ahead of the A100 both in terms of FP64 matrix and vector compute, while also offering a fair bit of additional memory and bandwidth. Add to that the fact that there won’t be any new offerings from NVIDIA in the data center market for at least another year, and AMD has got a substantial advantage for a substantial amount of time.

Areej

Computer Engineering dropout (3 years), writer, journalist, and amateur poet. I started my first technology blog, Techquila while in college to address my hardware passion. Although largely successful, it was a classic example of too many people trying out multiple different things but getting nothing done. Left in late 2019 and been working on Hardware Times ever since.
Back to top button