Originally Posted by Liranan
This started with the last Titan series, nVidia removed FP64 and people bought it anyway as it was a gaming card. Now they will release these Quadro's and people will still buy them and all because AMD use OCL and CUDA has this market in lock down.
Given that AMD didn't include FP64 on Fiji nor did Nvidia on GM200 and both were 600mm2
dies, my best guess is that both companies realized they were at TSMC's limits and could go no further. That seems to be the upper bound for any die, really, outside of some of Intel's >$5000 processors (such as the 62-core Xeon Phis that only ever had 61 enabled thanks to yields). Nvidia's Teslas
didn't even have a higher ratio. So that's kinda justifiable, and technically the Maxwell Titan, Quadro, and Tesla all had the same (crappy baseline 1/32) ratio.
As it stands, GP102 is supposedly 471mm2
while GP100 - with its HBM, FP64, and FP16 - is 610mm2
. I feel safe making the assumption that GP102 lacks all those features because that puts GP102 at 50% more than GP104 in all ways - cores, ROPs (probably), memory controllers, whatever - and the die size is exactly 50% larger than GP104. GP100's size isn't cheap, and it's 30% larger than GP102 for those features. On a given wafer, you can at best make 77% as many GP100s as you can GP102s. HBM vs GDDR5X really throws a spanner in my math, because I'm not sure about the size of the memory controllers. HBM moves some of it onto the DRAM chips, doesn't it? I wouldn't be surprised if a GDDR5X version of GP100 ended up closer to 700mm2
than 600.Edited by CynicalUnicorn - 7/25/16 at 6:14pm