06-02-2016, 03:39 AM
How does NV plan to milk the Pascal generation?
GP104: Just like Kepler GK104 (GTX 680) all over again - around 25% faster than previous flagship, much smaller and more efficient.
Then when AMD threatens with Vega that has HBM2 memory, Nvidia launches GP102 that has 50% more of everything (except for maybe slightly lower core clock due to its size - probably no more than 10% less). 40% faster is a safe bet.
A while after that, AMD tries its final attack with a Fury-like massive monster (up to 600mm^2). That is about as big as GP100, the big Pascal that Nvidia already announced (but probably won't be available for a long while). All this time, Nvidia has secretly been planing on GP110, with additional tweaks and refinements. It might be a bit like Maxwell vs Kepler, that largely did away with the FP64 double precision stuff. (I.E., TITAN BLACK, which is Kepler, has 1:3 double to single precision ratio, but then Titan X Maxwell has only 1:32 ratio.) GP100 has a full-fledged 1:2 ratio, which certainly does take up a lot of resources. It'd be like GTX 580 GF110 vs GTX 480 GF100 (which was never fully unlocked anyway, but you get the idea..), but maybe even better. I wouldn't be surprised if NV calls it GP200 instead of GP110.
Now, I'm putting my crystal ball back in the closet. Hope you guys enjoyed the show.
GP104: Just like Kepler GK104 (GTX 680) all over again - around 25% faster than previous flagship, much smaller and more efficient.
Then when AMD threatens with Vega that has HBM2 memory, Nvidia launches GP102 that has 50% more of everything (except for maybe slightly lower core clock due to its size - probably no more than 10% less). 40% faster is a safe bet.
A while after that, AMD tries its final attack with a Fury-like massive monster (up to 600mm^2). That is about as big as GP100, the big Pascal that Nvidia already announced (but probably won't be available for a long while). All this time, Nvidia has secretly been planing on GP110, with additional tweaks and refinements. It might be a bit like Maxwell vs Kepler, that largely did away with the FP64 double precision stuff. (I.E., TITAN BLACK, which is Kepler, has 1:3 double to single precision ratio, but then Titan X Maxwell has only 1:32 ratio.) GP100 has a full-fledged 1:2 ratio, which certainly does take up a lot of resources. It'd be like GTX 580 GF110 vs GTX 480 GF100 (which was never fully unlocked anyway, but you get the idea..), but maybe even better. I wouldn't be surprised if NV calls it GP200 instead of GP110.
Now, I'm putting my crystal ball back in the closet. Hope you guys enjoyed the show.