Ben Gilbert: belief

Transcript context

…Right. If you could solve that problem, approximately something like—I forget the number—70 milliseconds, 100 milliseconds, but it’s not that long. My point is, if you could remove the obstacles everywhere else, then the speed of light should be perfectly fine. You could build data centers as large as you like, and you could do amazing things. This little, tiny device that we use as a computer, or your TV as a computer, whatever computer, they can all instantly become amazing. That insight 15 years ago was a good one. Speaking of the speed of light—David’s begging me to go here—you totally saw that InfiniBand would be way more useful way sooner than anyone else realized. Acquiring Mellanox, I think you uniquely saw that this was required to train large language models, and you were super aggressive in acquiring that company. Why did you see that when no one else saw that? There were several reasons for that. First, if you want to be a data center company, building the processing chip isn’t the way to do it. A data center is distinguished from a desktop computer versus a cell phone, not by the processor in it. A desktop computer in a data center uses the same CPUs, uses the same GPUs, apparently. Very close. It’s not the processing chip that describes it, but it’s the networking of it, it’s the infrastructure of it. It’s how the computing is distributed, how security is provided, how networking is done, and so on and so forth. Those characteristics are associated with Melanox, not Nvidia. The day that I concluded that really Nvidia wants to build computers of the future, and computers of the future are going to be data centers, embodied in data centers, then if we want to be a data center–oriented company, then we really need to get into networking. That was one. The second thing is observation that, whereas cloud computing started in hyperscale, which is about taking commodity components, a lot of users, and virtualizing many users on top of one computer, AI is really about distributed computing, where one training job is orchestrated across millions of processors. It’s the inverse of hyperscale, almost. The way that you design a hyperscale computer with off-the-shelf commodity ethernet, which is just fine for Hadoop, it’s just fine for search queries, it’s just fine for all of those things—…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.