Coming from the horse's mouth directly (relevant time stamp 2:31 and also 4:50 ?t=151s ?t=290 )
And from the reviews here on the A750/A770 and looking at the performance graphs seems pretty in line with what Raja Koduri was describing, once you move the bottlenect to the gpu it's able to stretch it's legs more
Raja claims their driver is CPU bound, but this is nonsense for the following reasons;
If driver/CPU overhead were the bottleneck, then we would see a growing bottleneck with frame rate, meaning that once a A380 come close to this bottleneck, faster GPUs (of the same architecture) wouldn't be able to break through. But we do continue to see scaling with A750 and A770, so this is clearly not the case. As mentioned, synthetics and compute benchmarks scale better than games on this architecture, which is evidence of hardware level scheduling.
The fact that higher resolutions scale ever so slightly better conforms with hardware level scheduling too, as fewer bigger batches are easier to manage.
But keep in mind that higher resolutions doesn't unleash that claimed untamed power of this card, it's not like it's competing with RTX 3080 in 4K, so even though the driver overhead per time is reduced, there is barely any difference, because it's still bottlenecked by the hardware. So there is no way a better driver will suddenly unleash ~65% more performance to compete with a RTX 3080, as there is no evidence that the driver is locking up this performance.
And the common problem has a common origin - Raja Koduri
He is certainly a master of spin.
But the parallel with AMD is actually very relevant, and not just because of clown Raja. Back when Polaris released, it was a giant disappointment, but both the PR team and some reviewers led us to believe that RX 480 was just a diamond in the rough, needing just a
couple of months of driver optimizations and voila performance in the ~GTX 1070-1080 class. There were endless discussions about how much better of a long term investment it would be. And the spin continued with refresh (RX 500) and the Vega, the same nonsense about just give it a little time. And as we know now, these gains never materialized.
When AMD finally released the RDNA family, the Navi 1x cards performed a little better, even though their API was barely changed from Vega and Polaris and the drivers were basically the same (except for new hardware capabilities of course). RDNA2(Navi 2x) continued this trend, improving on the hardware side (especially resource management) while retaining the software pretty much the same, yet the old Polaris and Vega cards have still not gained any benefit from these new drivers.
So let this be a lesson for everyone; Driver maturity for lack of performance is a silly excuse. If they knew about a major potential in the drivers they would have fixed ahead of release, and it would be very evident as a bottleneck in the benchmarks.