Known Nvidia leaker kopite7kimi—who originally appear the name of the company's abutting GPU architectonics a few years ago—says Nvidia's next-gen abstracts centermost GB100 Blackwell alternation cartoon cards will accept a agnate amount calculation as Ada Lovelace chips but will accept "significant changes in its assemblage structure."
An beforehand tweet (via Hardware Info) indicates that the GB100 GPUs will use a chiplet-based multi-chip bore (MCM) instead of the accepted caked designs.
MCM is an avant-garde packaging address that splits GPU apparatus beyond abstracted dies. The leaker doesn't acquaint us what the agreement will attending like—such as whether it will accommodate assorted GPU compute dies, or aloof afar anamnesis dies alongside a distinct GPU core—nor does it advance whether the GB100 will be the alone Blackwell GPU with a multi-die design.
This about-face from a caked focus would accord Nvidia's abstracts centermost GPUs a lot of jerk allowance for abeyant dent customization in adjustment to attempt with AMD's Aptitude MI300 alternation chips. The red team's own abstracts centermost chips are abstraction a whole Zen 4 CPU central an actually massive GPU.
AMD additionally already has some chiplet gaming GPUs that we like so far, such as the AMD Radeon RX 7900 XTX, admitting its countdown chips still don't accept assorted compute dies, aloof abstracted compute and I/O and anamnesis dies.
After the dramas of GA100 and GH100, it seems that GB100 is assuredly activity to use MCM.September 18, 2023
Anyways, the chips rumoured actuality are accurately abstracts centermost GPUs, which ability not be cool absorbing to us PC gamers. Admitting the Blackwell architectonics is reportedly advancing to our GeForce cards, beneath the GB200 name. Still, if Nvidia has begin a way to accomplish multi-compute dies action aural a acceptable bold apprehension activity again we ability still accept article actuality to get actual aflame about.
Because that's the catchy thing. Back it comes to abstracts centermost GPUs, accomplishing a accomplished abundance of austere cardinal crunching, rather than acceptable anatomy rendering, is a lot easier to do back advance beyond assorted compute dies. That's why SLI and Battery multi-GPU band-aid kept on activity in spirit in the compute amplitude continued afterwards it was arbitrarily accomplished in the gaming space.
That's about what we're attractive at back it comes to multi-compute MCM cartoon cards: an SLI/Crossfire band-aid active into a distinct chip. And there's accurately no acknowledgment actuality of whether Nvidia is attractive to cycle that out to its next-gen GeForce cards.
We can be hopeful, but we'd apparently advance it's absurd in this abutting GPU generation. But it's been accounted the abutting bearing of Nvidia gaming GPUs won't appear until at atomic 2025—plenty of time to amount that all out, appropriate Jensen?
