Short answer: Citing an article, Sam Nadler says the Rubin architecture runs three and a half times faster than Blackwell on model training and five times faster on inference. Jordan Metzner says each Rubin has 72 GPUs.
Watch this answer from Episode 26 (starts at 25:49) · Episode 26 notes and transcript
From Episode 26: We Built an AI Trading Bot for Prediction Markets — Here’s What Actually Happened, starting at 25:49.
Sam Nadler [25:48]: But, yeah, NVIDIA launches the new Rubin chip, and I know you have a bit more insight on why this chip is actually different, but it seems like it's like the next wave of or will enable the next wave of massive AI models.
Jordan Metzner [26:03]: Yeah. So, I mean, look. Obviously, they being NVIDIA is trying to, consistently improve the technology of their chips. They are obviously building, like, better packs and better formats and new implementations. This new chipset is called a Rubin or a Rubin Pod. And I think each side inside each Rubin is 72 GPUs, and then I think inside of each GPU, there's more than a thousand plus chips. So we're talking like thousands and thousands of GPUs at the core that are able to power these things. And, as the market leader, we're gonna continue to see new use cases that were not available previously. So, we're gonna see increased inefficiency. We're gonna see increase in new applications, and hopefully some new discoveries, whether that's scientific, mathematical, or even things that are fun like video generation, image generation, self driving, and whatnot. So overall, it just seems like NVIDIA is kind of, essentially kind of they're striking on all cylinders. But, yeah, it's you know, I mean, I did not expect this coming out of NVIDIA at the first of the year. And, yeah, the game's heating up and, competition is probably good for the market, and hopefully, this will continue to put pressure on Tesla to continue to improve their product.
Sam Nadler [27:14]: Yeah. Just to be specific, it's from this article, it says the Rubin architecture will operate three and a half times faster than the Blackwell architecture on model training and five times faster on inference tasks. I mean, that's that's a pretty big increase in performance. So, I think to me, obviously, it was expected that I assumed, I guess, for me that NVIDIA would keep pushing the boundaries on new chip development, but I think it's really exciting. They're entering into this autonomous driving space. Super bullish about the next several years with NVIDIA. And, yeah, anything else in the news or NVIDIA? Anything else you wanna cover?
