We’re talking about a rumor here, of course, and all this is to be taken with a pinch of salt. However, it comes from @kopite7kimi on X.com, a leaker who’s proved his worth… Enough to be listened to. And according to him, the GB202 gaming chip uses the same manufacturing process as the B100 server chip.
GB202: a chip engraved using the TSMC 4NP process?
If we are to believe what is being said, the potential future RTX 5090 could be etched using the same process as the B100. However, the leaker indicates that this would be 4NP, which he doesn’t find in TSMC’s nomenclature. Speculation is therefore rife that this could be a customized version for NVIDIA, based on the TSMC 5 process. However, this would clearly benefit the card, since it enables a 30% increase in transistor density.
GB202 will use the same process node as GB100. I must clarify once again that TSMC 4N(vidia) is based on TSMC 5, not 4nm.
I'm sorry I cannot match Jensen's naming with TSMC's naming. We need professional chip analysis to determine.
At least, there is a 30% increase in density.— kopite7kimi (@kopite7kimi) March 19, 2024
Another area of improvement for the GPU is the L1 cache, which could be revisited. On the Ada Lovelace and Ampere GPUs, we found 128 KB of cache. This amount could be increased, improving SM throughput.
https://twitter.com/kopite7kimi/status/1769898435367620933
Finally, as far as the GB202 is concerned, previous rumours suggested a GPU with 192 SM, i.e. a potential 24,576 cuda cores if the number per SM remains unchanged. As for memory, we’re talking about a 512-bit bus with GDDR7 memory running at 28 Gbps.
Last but not least, the chip could also benefit from an MCM design with two chiplets, rather like the B200 we mentioned earlier this morning. This would then come in very handy for producing RTX 5080s, possibly based on a GB203 that would be half a GB202 (96 SM memory in 256 bits).
