Originally Posted by
DanceswithUnix
But the 2080 cards still have an fp32 focus. Volta could do fp64 at half the rate it could do fp32, like you expect from a commercial focused card. The 2080 can do fp64 at 1/32 of the fp32 rate, like you expect from a consumer card.
Whilst I'm sure the tensor cores will be what they developed for commercial users, their inclusion does not make it a commercial card. It just means Nvidia think the feature is worth the sacrifice in silicon area over putting more shaders in. Then there is the ray tracing support; is there any support for that in commercial render engines? Something that gives an iffy quality lighting system good enough for action games isn't likely to impress the likes of Pixar in rendering their latest movie where every pixel should be spot on.
So AFAICS this is a consumer part, probably a Volta with the FP64 stripped out and some raytrace tech added giving a slightly smaller die that GV100. Given Volta and Turing are both 12nm products, I wonder if Nvidia have done exactly the same commercial/consumer split as before they just staggered the release.
As an aside, there was a die shot of a Turing compute unit that implied a quarter of the area was for tensor cores and a quarter for RT, so they could have had twice the shaders if they cut those out and scaled up the number of CMs to fill the space. I have to wonder what that would do the things like the anti-alias performance if it could generate sample spots at twice the throughput.