There’s the suggestion going round that the inexperienced group is testing out a chiplet model of its high GeForce graphics card for the longer term Nvidia Blackwell GPU technology. This comes scorching on the heels of one other hearsay stating explicitly that the Blackwell architecture is going to be used for data center silicon, too, and that this successor to Hopper will certainly be a multi-chip module GPU.
We’re at least a year out from the release of the Nvidia RTX 5090, however that is not going to cease the hearsay mill going into overdrive each time there’s the faintest whiff of next-gen silicon within the air. And meaning we’ve got to take each hearsay with the requisite measure of salt on condition that at this level it is all largely guesswork, conjecture, and rumour.
The thought of Nvidia lastly making the change over to a chiplet design for its server-level GPUs would not shock, and certainly feels like a sensible plan. The kind of compute-based workloads vital in knowledge facilities will be pretty simple to run throughout a number of graphics playing cards, so the concept that they might run on a number of compute chiplets inside a single GPU package deal would make sense.
There are additionally some guessed-at potential RTX 5090 specs floating about, however they’re so skinny as to be nearly clear. The actually intriguing query is round whether or not Nvidia could make GPU chiplets work within the gamer area. AMD has already nominally made the transfer to chiplets for its personal high-end graphics playing cards, however the Navi 31 and Navi 32 chips solely utilise a single compute die (the GCD) inside the package deal, so video games successfully nonetheless solely have to have a look at one GPU.
There are reviews this can change with subsequent RDNA 4 chips, although, with rumours claiming AMD is going to abandon chiplets and even the high-end market with its next-gen GPUs.
There is not any trace within the tweet by the XpeaGPU account—itself in response to the Kopite7Kimi tweets about Blackwell—that the GB101 chiplet model it is referencing comprises a number of compute dies, simply that Nvidia has such a GPU “underneath investigation.” The actual fact it is not permitted but means that the success of the endeavour continues to be to be verified.
That is likely to be as a result of it is making an attempt one thing formidable, like a number of compute dies for sport rendering, or just because it’s doing extra testing. Whereas it is not specific within the tweet that they’re speaking a few GeForce card, similar to a possible RTX 5090, the reference to Ada Subsequent suggests the chiplet model can be for a successor to the present gaming GPU structure.
The the Ada Lovelace structure has additionally been utilized in skilled playing cards, which themselves are extra compute-focused than rendering, and a professional chiplet card cannot be dominated out. Nor can a resurrected Titan Blackwell card, y’know only for the folks with more cash than sense. They will not care if it really works correctly or not…
We should still find yourself within the scenario the place all of Nvidia’s Blackwell GPUs within the GeForce RTX 50-series are monolithic chips—which I feel might be the almost certainly consequence—and the suggestion is that the entire lower-end consumer GPUs might be anyway. However the concept the inexperienced group is a minimum of testing it out is tantalising.
As a result of the way forward for GPUs must be chiplets, whether or not that is the following technology or in a number of generations time. The highest-end monolithic graphics silicon is getting so large that they are monstrously costly to fabricate, and disaggregating GPU elements into discrete chiplets cuts that down vastly. It additionally successfully means you may develop a single compute die, the actually advanced GPU logic chiplet, and use multiples of it to create all of the completely different playing cards in a stack. Much less R&D prices for various GPUs, decrease manufacturing prices, win and win.
It is simply that rendering sport frames throughout a number of chiplets is hard.
You solely have to have a look at the mess SLI and Crossfire ended up in making an attempt to try this throughout a number of graphics playing cards to see how troublesome it’s. However not unimaginable. The primary chip maker to get a gaming GPU operating fortunately with a number of compute dies is sort of actually going to scrub up.
