| hinch said: Yeah we can see so far at how moving to TMSC's 4nm has added a lot of efficiency. Even with like for like specs on Ada. With the performance gained quite substantial from higher boosting clocks and massively increased cache. There's probably already number crunched estimates already to see what it performs like (some outlets say 3080-like performance) which isn't too bad of an uplift all considering. I'd be interested to see how the 4070 stacks up vs the 3070 like the same frequency. Then we can see the IPC gains are like this gen. I'm just really drawn by the incredibly low power consumption of this particular SKU. Lets hope the FE doesn't sell out within 5 minutes lol. |
Yeah, the 4070 is probably to the 3080 what the 4070 Ti is to the 3090 Ti, so a nice performance per watt uplift overall.
I suspect this is as much on the new node as it is on Ampere having been a relatively poor leap in the first place (for any application not using sparsity instructions or the added FP32 units). Nvidia was once twice as efficient as AMD following Maxwell and Pascal, and their lead had evaporated completely in just four years. This seems too much to ascribe just to RDNA.
