Nvidia trademarks "Aerial" and "Hopper": New MCM GPUs on the way?
Nvidia recently trademarked two items at the US Patent and Trademark Office: Nvidia Hopper and Nvidia Aerial. The former, Hopper, is especially interesting: Rumors earlier in the year indicated that Hopper was the codename for an upcoming MCM (multi-chip module)-based GPU architecture.
Last June, tipster @kopite7kimi claimed that Hopper will utilize MCM to "build giant cores." At the time, there was no corroborating evidence to back the "Hopper" claim. Nvidia's new trademarks change the situation.
At this point, there is precious little we know about Hopper apart from the name itself and @kopite7kimi's claim that it will be an MCM architecture. Why is called Hopper? And why does going the MCM route matter?
Hopper is evidently named after Grace Hopper. During World War II, Hopper was a computer programmer for the Harvard Mark I. The Mark I was an electromechanical computer that helped with, among other things, the development of the US nuclear program. Hopper made significant early contributions to the development of computing: the COBOL langauge is based on work she did on FLOW-MATIC.
What about Hopper being an MCM architecture? With Moore's Law slowing down, MCM is being explored by AMD, Intel, and Nvidia as a way to deliver year-on-year performance improvements and scaling. MCM is what, for example, makes the 16 core Ryzen 9 3950X possible at reasonable prices.
With manufacturers struggling to shrink transistors below the 7nm point, Hopper may be a key inflection point in Nvidia's strategy. Multiple high-end dies in a single MCM package could return us to the days where GPU performance doubled every few years. If this is the case, playable 4K experiences might be possible on PC without a $1400 RTX 2080 Ti.