Originally Posted by The Stilt
While GP106 is 13.8% smaller (200 & 232mm²) than Ellesmere, it is also clocked ~35% higher (1709 & 1266MHz). Most importantly the performance per watt on full Pascals (GTX 1060 & GTX 1080) is 53 - 79% higher than on RX 480. The fact that the 79% higher performance per watt over RX 480 is achieved on a > 35% larger and 37% higher clocked GPU is simply amazing.
I'm not saying that the RX 480 would hit anywhere similar performance per watt or frequency figures if made on TSMC 16nm FF+ instead of 14nm LPP, however most likely both of these figures would be significantly
better if it was.
AMD's GCN architecture is just not very energy efficient.
My R9 290 pulls about 420W at 1.15GHz. I have 2560 SPs, versus RX 480's 2304,twice as many memory chips, and more ACEs of course, which explains away about 30W of power usage improvement right off the bat.
So, at 1.15Ghz, 2304 SPs, and a 256-bit bus with eight RAM chips, about 390W should be expected... at 1.15Ghz. RX 480 pulls ~165W at about 1.2Ghz average. In addition, RX 480 includes more cache and other improvements not generally good for reducing power usage... more like using more.
So 14nm LPP looks to have reduced power consumption by well more than half (only using 42% of the power), while also allowing for higher clock-speeds on a historically low-clocking design. What's more is that RX 480 power usage doesn't climb severely with increased clocks during overclocking compared to 28nm GPUs.
On the GTX 1060 front, we can only compare to Maxwell 1280 cores... which is 77% of what is found on the GTX 970. It has a third fewer memory chips as well. So, right off the bat, we should expect cut-down 145W GTX 970 to pull only 110~120W... without any process improvements at all. GTX 970 pulls ~145W at 1178MHz, but would only pull 120W in the GTX 1060 configuration. In order for this new GPU to match the old GTX 970, it would need to overclock to 1.566GHz. And it does. And more - all the way to 2Ghz... because nVidia put in significant effort in increasing clock-rates on Pascal. Of course, higher clocks means more power usage... but we know what the GTX 970 pulls at higher clocks - because it can reach them. At 1.25Ghz, the 970 uses about 160W - about 7W more than you'd get from linear scaling. The curve is established with 1.4Ghz pulling ~190W. At 1.84Ghz, if possible, I'd expect the card to pull 265W. *However, this doesn't include the reduction from the smaller GPU - so we should call that 240W.
GTX 1060 pulls 120W, nearly exactly, on 16FF+, at 1.838Ghz. That is a 2.21x improvement for 16FF+, compared to a 2.36x improvement for 14nm LPP. This is slightly less of a good showing for 14nm LPP than expected, but that is most likely due to variation in the design (larger caches, VRM efficiency, and so on), and the fact that I didn't correct the 50Mhz difference between the R9 290 figures I have and the RX 480 figures...
Expecting a 240W draw and getting a 120W draw, 16FF+ provided 2:1 improvement, a worse showing than 14nm LPP by a fair bit.
If GTX 1060 was on 14nm LPP, it'd draw 15~20W less!
14nm LPP's only issue right now is in its very high leakage variance between parts of the wafer - something that may well improve over time... or better binning could effectively hide.
Also, I did all of this in my rush to go to bed, so I did 90% of the math in my head and some of the numbers are from memory (but I have a thing for numbers...), so I hope things are accurate, correct me if I'm wrong
Well, it took me all of a few seconds to realize I didn't scale-down the power usage from the GTX 970 levels prior to calculating 16FF+ power savings, so I fixed that...