MistaSparkul
2[H]4U
- Joined
- Jul 5, 2012
- Messages
- 3,524
88.8% of AD102 (4090) vs 97.6% of GA102 (3090), on similar memory bandwith with DDR7 not ready:
Around 65% faster, in the low end of that 60%-80% expectation.
Lovelace power-core scaling is not that good, but a 97.6% activated AD102 would probably reach that 70% higher performance, Lovelace ended up not pushed at all and with no memory bandiwth increase, lot of product getting the 18176 core good chips sold at much higher pricing than the 4090.
Was it designed to be a 18xxx core-550watt model at some point and time that looked to be 60-80% above the 3090-3090TI, maybe.
The 380mm lovelace (4080) ended up easily higher than that 60%-80% boost from the 392mm Ampere.
Unlike Ampere->Lovelace, I doubt they will achieve this kind of boost while reducing the die and memory bus, the boost solely from how much better of a node they upgrade too should be much smaller this time.
Reading now the idea that they would have any issue reaching RDNA 3 raster performance.... they were able to pretty much do it with a "4070"
Just read the 2nd point on the slide about RDNA3 lmfao.