In contrast to its rivals AMD and Intel, Nvidia but has to implement a multi-chiplet design for high-performance compute GPUs, but it surely seems to be like the corporate is likely to be on observe to lastly use this design format with its next-generation Blackwell era of GPUs. Not less than so says famend {hardware} leaker @kopite7kimi, who tends to have correct details about Nvidia’s intentions. For now, that is all a pure hypothesis although.
“After the dramas of GA100 and GH100, plainly GB100 is lastly going to make use of MCM,” @kopite7kimi wrote in an X submit. “Possibly GB100=2*GB102.”
Nvidia’s GA100 and GH100 compute GPUs characteristic a die measurement 826 mm² and 814 mm², which may be very near the utmost reticle measurement of 858 mm2. Producing such giant chips with good yields is difficult, but it surely seems to be like TSMC does it fairly properly since Nvidia actually ships tons of H100 and A100 compute GPUs each quarter.
Whereas Nvidia has constantly managed to extend efficiency of its compute GPUs fairly tangibly with each new era to date, reticle measurement remains to be a difficulty. Utilizing a multi-chiplet design would allow Nvidia so as to add extra transistors into its next-generation compute GPU and enhance efficiency achieve over its H100 extra considerably than it might do with architectural enhancements alone.
Since each AMD and Intel have already adopted multi-chiplet designs for his or her compute GPUs and can solely additional enhance chiplet and transistor depend going ahead, Nvidia may not have a alternative however to embrace a multi-tile design too. It stays to be seen whether or not the corporate would go together with a dual-chiplet (like AMD’s Intuition MI250) or multi-chiplet (like AMD’s Intuition MI300 or Intel’s Ponte Vecchio) strategy, however the firm simply can’t ignore benefits introduced by trendy packaging applied sciences.
If we proceed to invest, we’d assume that Nvidia will undertake multi-tile design just for its Blackwell GPUs for AI and HPC compute, whereas its gaming GPUs will stay monolithic. This would possibly make sense since making two GPUs work in parallel is difficult. However then once more, Nvidia may not be capable to ignore multi-chiplet designs even for shopper PCs within the Excessive-NA period since ASML’s next-generation scanners will halve reticle measurement (to 429 mm2) and Nvidia won’t be able to handle high-end gaming machines with monstrous monolithic GPUs just like the AD102 (609 mm2) except they use a minimum of two chiplets.