RDNA3 and Ada, Ampere are pretty similar in that they doubled FP32 units, while only half is capable of executing also INT32.
Yet, Ampere saw a lot bigger IPC(PPC) gain than RDNA3. Why? Is this truly only driver fault or Nvidia is using more HW.
Strangely enough, from my findings, compute uplift and meta launch day gaming performance uplift (3Dcenter) shows almost identical gains for both Turing to Ampere and RDNA2 to RDNA3:
3090 to 2080Ti and 7900XTX to 6900XT:
~2,65x compute and ~47% raster 4K both, ~58% and ~61% 4K RT 3090 and 7900XTX.
Bare in mind both are competing with their previous gen w/ pristine drivers. Versus launch day 2080Ti and 6900XT performance would be different, but that's beyond the scope of my comment.
The main difference is RDNA3 arguably saw a bigger node jump from RDNA2 than than Ampere saw from Turing. And AMD clearly planned for bigger gains. Hopefully it's not all down to hardware mishaps and they can improve the situations by drivers.
I compared boost full N23 vs boost full N33, because I thought he was talking about boost clock.
We still don't know If he was talking about boost or game clock, so you can't confidently say he was talking about game clock. With this in mind, you can't really say I was comparing apples to oranges.
We don't even know If It's 2700MHz, he said comparable to RX 6650XT AIB and I just guessed this frequency.
I will compare full N23 vs full N33:
Boost clock: 2416MHz vs 2700MHz (+12%)
Game clock: 2162MHz vs 2700MHz (+25%)
If I consider that he said desktop RX 6750XT level of performance on average, then this clock really should be game clock, It would be just a bit higher(+8%) than 2495MHz for RX 6750XT.
The rest has to come from the architecture ad that would be ~20%.
That's a lot.
Yeah that's what it came down to for me, what clocks are we comparing.
I wonder if desktop N33 is the same silicon as these soon the be announced laptop chips or newer. And how much further could a 150-180W TGP SKU clock from the specs we'll see at CES.