GA103-200 22bn, TU102-300 18.6bn.
I get that but my point is that size (transistor count) isn't really related to performance.
40 series cards are 25-30 percent cache that is 3-5 times denser in transistor use.. so a large percentage of the die transistor count is in the cache with 40 series. Also 40 series has not really changed much from Ampere at the SM level just a few beefed up tensor cores..
I need to look at the white paper for the 40 series against the 30 series again, but from what I remember the only real big changes were smaller manufacturing process and more SM's /cores. Example the 4090 has about 60% more cores than a 3090 and a lot more cache 16 times more.. But is 55-60% on average faster than a 3090.
A lot of them transistors are in that cache on 40 series and need to count that in too.. Reality is ADA is Ampere shrunk with more cache and faster clocks because of the new manufacturing process and high power use still 450w on a 4090 is not in real terms amazing power use... but yes we can say it is efficient compared to previous gens when comparing its performance, but reality is we need to add more and more transistors and more and more power now to see benefits in performance with current gpu architectures.
Last edited: