That's why I said even If.
It's true, that RDNA2 low and mid-range GPUs had a significant
I don't know how you got 2.61GHz for the stock XTX during 3DMark.
It's pretty suspicious that a 12% increase in clockspeed results in 16% increase in performance when BW increased by only 10%.
View attachment 80758
If you compare max OC RX 7900 XTX vs RX 7900 XTX Taichi then by increasing clockspeed by 8% you gain only 4.4% when the BW is the same.
This would mean 22gbps is enough for ~2835MHz. Now, If I divide this clockspeed with the increase in performance, I end up with 2835/1.16=
2444 MHz In case the scaling is linear.
The 2.61 comes from the efficiency page where he gives you the average clock achieved across all the titles. Not ideal but what we have. Entirely possible the stock XTX is clocking lower than this in 3DMark. Alternatively it is possible the 94FPS result for the stock XTX is with an older driver or some other system level difference. It depends if he retests or not.
At 2.8Ghz ram the 2795 Taichi vs the 2853 TUF shows linear scaling. 2.1% clockspeed increase vs a 2.2% performance increase.
Even the 2921 Merc is getting a 1.5% performance increase from a 2.4% clockspeed increase and a 1% bandwidth regression.
Also if you compare the Merc to the Taichi white you get a 0.9% performance increase from a 1.1% bandwidth increase and a 0.5% clock speed regression.
Also based on the 2.61Ghz average for the stock XTX vs the 2.73Ghz average for the Taichi white that is a 4.6% clockspeed increase and the 4K performance ranking has the Taichi White 4.2% faster than the stock XTX.
Even the 7900XT to 7900XTX is pretty linear in that table if you factor in the extra bandwidth the XTX gets.
So it does look like RDNA3 scales in a pretty linear fashion provided there is enough bandwidth upto what seems like the 3Ghz region at minimum.
If we do some rough math then 3Ghz sustained N32 would have a 12% clockspeed advantage over 7900XT but only 71% of the shaders netting out to around 80% of the performance. OTOH a 70CU N31 with 2.69Ghz sustained like the 7900XT manages would have equal clocks and 81.4% of the shaders.
80% of 7900XT is around 6900XT tier performance. So I still stand by my thinking that if N32 can clock to 3Ghz + at a sane power envelope AMD will make that the 7800XT. If they cannot get it there they will probably make the 70CU N31 the 7800XT. Either way you are looking at 6900XT to 6950XT performance at 4K by the looks of it.
Now I am aware this is over simplified because with 3SE vs 6SEs in the 7900XT there is going to be a substantial ROP deficit which might make scaling to 4K worse. That is where the 5SE 70CU N31 design would be better because it won't have the same ROP deficit making it a better 4K card even with equal VRAM and bandwidth.