Well let's assume it is completely genuine and also assume that Time Spy is completely compute/shader bound (it's not, but you have to start somewhere). The purported Ampere is clocked to 1935 MHz, whereas my overclocked 2080 Super averaged a core frequency of 2030 MHz during the tests - so the new chip was 4.7% lower, but achieved a graphics score 46% better.
So clock-for-clock, it's roughly 50% better. The TU104 in the 2080 Super has 6 GPCs, each with 4 TPCs, and each of those with 2 SMs, giving a total of 3072 CUDA cores. Sticking with the compute bound assumption, the Ampere chip would need to have 4608 cores to reach that 50% improvement - the same number as in the TU102.
I've
taken an educated guess at the full consumer GA102 being a 6 to 8 GPC, retaining the 6:1 TPC ratio (same as Turing, same as the GA100). This would put the CUDA core count from 4608 up to 6144 - like this:
6 GPCs = 4608 CUDA cores > '3080'
7 GPCs = 5376 CUDA cores > '3080 Ti?'
8 GPCs = 6144 CUDA cores > '3090 / Titan A?'
My 2080 Super is
roughly 18% better in Time Spy (even accounting for clock differences) than my previous Titan X Pascal, so if this is a 3080, then it would follow a similar pattern of generational increments - in that the xx80 model is as good or better than the previous top-end version.
Given that the GA100 is only 1.3% larger (die area) than the GV100 it has replaced, Nvidia/TSMC are clearly happy to be churning out very large chips again, so there's a good chance that a full GA102 is an 8 GPC chip.