Nvidia’s ever-optimistic CEO, Jensen Huang, has revealed extra particulars of the corporate’s next-gen AI platform on the identical GTC occasion the place the corporate’s quantum computing plans had been outlined. He calls it the Vera Rubin “superchip” and claims it incorporates no fewer than six trillion transistors. Yup, six trillion.
In quite simple phrases, meaning the Vera Rubin superchip sports activities 60 occasions the variety of transistors of an Nvidia RTX 5090 gaming GPU. And that could be a utterly bananas comparability.
Associated articles
N3 is extra dense than N5, for positive. However not something like sufficient to go from about 100 billion to 6 trillion transistors. Truly, it seems that Vera Rubin is not simply multiple chip, it is multiple chip package deal.
Huang held up an engineering pattern of Vera Rubin and what’s clear to see are the three essential chip packages, one for the CPU and two GPU packages. In a later a part of the presentation, it turns into clear that every GPU package deal truly incorporates two GPUs. In order that’s 4 GPUs complete, plus the CPU. An in depth up on the Vera CPU package deal additionally reveals it to be a chiplet design, additional elevating the general transistor depend headroom.
That stated, every Rubin AI GPU die is monolithic and described as reticle-sized, which implies it is as huge as TSMC’s manufacturing course of will enable. However even should you had been beneficiant and assumed that TSMC may squeeze 200 billion transistors right into a single monolithic chip utilizing N3, you’d want 30 such chips to get to that six trillion determine.
Truly, Nvidia additionally revealed a “Rubin Extremely” GPU package deal with 4 Rubin GPU dies, so these items is simply going to scale and scale. However, nonetheless, it is unclear how Nvidia is doing its transistor depend. I would need to assume that the whole lot, together with all of the reminiscence for the GPUs and CPU, plus the supporting chips on the Vera Rubin motherboard, is thrown in.
Nevertheless Nvidia is arising with the numbers, the Vera Rubin AI board (versus “chip” as Huang is looking it), is an unfathomably advanced factor. Certainly, Huang says it is 100 occasions sooner than the Nvidia equal product from 9 years in the past, the DGX One, which Huang says he delivered to what should then have been a nascent OpenAI.
The remainder, as they are saying, is historical past. And trillions and trillions of transistors. And trillions of {dollars} for Nvidia. For the document, Huang says Vera Rubin goes into manufacturing “this time subsequent 12 months, possibly even barely earlier” and it in all probability will not be low cost. So get saving up.
Finest mini PC 2025
All our present suggestions
