Nvidia has historically refrained from adopting a multi-chiplet design for its high-end compute GPUs. However, emerging information suggests that Nvidia may be contemplating this approach with its upcoming Blackwell GPU generation. This assertion stems from reputed hardware analyst @kopite7kimi, although it's crucial to treat such information as speculative currently.
A defining characteristic of the anticipated Blackwell GPUs is the reference to a Multi-Chip Module (MCM) design for the GB100 data-center GPU. Through this advanced packaging method, distinct GPU components can reside on separate dies, potentially offering Nvidia enhanced chip customization flexibility. Such a move might position Nvidia to adapt its chips more adeptly for both consumer and enterprise-specific requirements, possibly setting them apart from competitors like AMD.
The eventual release of Blackwell may mark a significant divergence from Nvidia's traditional GPU design methodology. For context, AMD's forthcoming MI300 GPU, which integrates chiplets, emerges as a likely rival. The advantage of chiplets is clear: they facilitate smoother integration due to smaller die sizes, enhancing wafer yields and paving the way for economical chiplet usage.
"After the dramas of GA100 and GH100, it seems that GB100 is finally going to use MCM," @kopite7kimi wrote in an X post. "Maybe GB100=2*GB102."
The substantial die sizes of Nvidia's GA100 and GH100 compute GPUs, at 826 mm² and 814 mm² respectively, approach the maximal reticle size of 858 mm². Consistently manufacturing such expansive chips at high yields is no small feat. Yet, Nvidia has effectively achieved this, as demonstrated by the significant shipments of H100 and A100 compute GPUs quarterly. Despite Nvidia's continual advancements in compute GPU performance, the challenge of reticle size remains. Transitioning to a multi-chiplet design could facilitate the incorporation of a higher transistor count in Nvidia's next-gen compute GPU, leading to notable performance gains beyond mere architectural improvements.
From a speculative standpoint, Nvidia might reserve a multi-tile configuration exclusively for Blackwell GPUs targeting Without such a transition, Nvidia could face challenges serving high-end gaming platforms with expansive monolithic GPUs, like the AD102 (609 mm²).