Still, everyone's products so far on 14/16nm processes have had a nice increase in frequency at similar power consumption. Apple's A9 and A9X (was it 400-600MHz on a constrained environment?), GP104 and that 2.1GHz shot and this rumor that custom cards might do 2.3-2.4GHz (so, around 0.6-1GHz up from GM204? If I may, these 2.3-2.4GHz GP104 cards are probably going to be consuming north of 225-250w, it sounds like they're pushing the chip as far as it'll reasonably go)
No reason not to think that P11 in that demo was running conservative clock speeds as with like any other usual engineering sample be it CPUs or GPUs, or to emphasise perf/w. AMD's video there states 850MHz at 0.8375v, that could do wonders in a notebook... which is its intended place. P10 will use higher voltage and higher speeds.
From the new process alone you have this inherent clockspeed increase that might put GCN at around 1.3-1.6GHz comfortably, if we keep 28nm's GCN perf/power curve. GCN doesn't clock as high as Maxwell, that's clear, yet it doesn't need such high clockspeeds to be competitive or outright beat or embarrass nV's competition in >2015 games. This is particularly clear of Hawaii vs GM204, Fiji as imbalanced as it is can't overcome GM200 this way. Tonga and the dinosaur Pitcairn do alright for what they are vs GM206. Just think of what a hypothetical >1.5GHz Hawaii could do, that is reference GM200 performance right there, if this were just a straight shrink.
Obviously GCN's power/perf curve is going to change, most of the architecture's critical parts are "new", and this time CUs are included (I remember reading GCN1-3 CUs are mostly the same in B3D, the rest of the GPU is what's been changed), so yeah, not much more to go by at this point. P10 might punch way above its weight with just 2304-2560SPs, just like a >2GHz 1xxx-2560SPs GP104 will just because of sheer clockspeeds, taking into account the hardware both chips have vs their 28nm 600mm^2 relatives. This is a node shrink after all, last time we had one of those was back in 2011-2012.
Computex can't come soon enough.
No reason not to think that P11 in that demo was running conservative clock speeds as with like any other usual engineering sample be it CPUs or GPUs, or to emphasise perf/w. AMD's video there states 850MHz at 0.8375v, that could do wonders in a notebook... which is its intended place. P10 will use higher voltage and higher speeds.
From the new process alone you have this inherent clockspeed increase that might put GCN at around 1.3-1.6GHz comfortably, if we keep 28nm's GCN perf/power curve. GCN doesn't clock as high as Maxwell, that's clear, yet it doesn't need such high clockspeeds to be competitive or outright beat or embarrass nV's competition in >2015 games. This is particularly clear of Hawaii vs GM204, Fiji as imbalanced as it is can't overcome GM200 this way. Tonga and the dinosaur Pitcairn do alright for what they are vs GM206. Just think of what a hypothetical >1.5GHz Hawaii could do, that is reference GM200 performance right there, if this were just a straight shrink.
Obviously GCN's power/perf curve is going to change, most of the architecture's critical parts are "new", and this time CUs are included (I remember reading GCN1-3 CUs are mostly the same in B3D, the rest of the GPU is what's been changed), so yeah, not much more to go by at this point. P10 might punch way above its weight with just 2304-2560SPs, just like a >2GHz 1xxx-2560SPs GP104 will just because of sheer clockspeeds, taking into account the hardware both chips have vs their 28nm 600mm^2 relatives. This is a node shrink after all, last time we had one of those was back in 2011-2012.
Computex can't come soon enough.
Last edited: