Regarding the chiplet architecture, for example, it looks like the GB102 chip for GeForce will continue to be a monolithic chip and not chiplet like AMD. However, this does not mean that we will not encounter chiplets at all. This architecture could appear within Blackwell for HPC chips (but probably not for standard GeForce). As for the manufacturing process, it is still expected to use TSMC’s 3nm process, which is 25% more expensive per wafer than the 5nm/4nm process. The question is how the chip area will change with the transition to a smaller process. A better process should mean a smaller chip, but the number of transistors will certainly increase again, which, on the contrary, increases the area.
It probably won’t surprise anyone that a frequency of over 3 GHz is expected, e.g. 96 MB of cache memory is also mentioned. Overall performance should double against the Ada Lovelace architecture, and the unanswered question is how the new cards will fare. Let’s hope that the manufacturer succeeds at least as well as with Ada Lovelace, which offers much higher performance for quite interesting consumption (for cheaper cards, consumption should even decrease between generations).
