Cuda vs Cuda

[jjohnthedon1]

Are there any differance in the different generations of cards in terms of Cuda performance

say a Kepler gpu with 1000 Cuda cores at 1000mhz 

vs pascal gpu 1000 Cuda cores at 1000mhz 

 

have the the cores got better ipc or are they the same ? 

[ChrisCross]

Just now, jjohnthedon1 said:

 

normally from jumping to a smaller architecture you get better ipc but im not sure. didnt see cuda or render benchmarks so far. double precision would be interesting too.

[GER_T4IGA]

6 minutes ago, ChrisCross said:

normally from jumping to a smaller architecture you get better ipc but im not sure. didnt see cuda or render benchmarks so far. double precision would be interesting too.

Nvidia claims a way higher double precision power compared to the previous Maxwell architecture which was even compared to the one before (kepler) pretty weak. I don't have any sources or quotes on this but from what I recall Kepler dropped to like 1/4th of the normal performance but Maxell to a 1/32th or something. Pascal was roumoured, teaserd or whatever (I don't remember where i read this or heard it - sry) to almost hit half of the normal performance but we haven't seen any benches as of yet just like ChrisCross said.

[Naeaes]

I have no basis for any of this but I'll speculate. It'd make sense that a smaller manufacturing node alone would give better performance for CUDA like it does for everything else. But AFAIK there's no common benchmark to make apples to apples comparisons, like that, enabling exactly 1000 cores at 1GHz and see which does better. It would be very interesting though. In this article on Wikipedia, they list CUDA versions from 1.0 to 6.1, their capabilities and respective GPUs. I barely understand any of it, I can imagine there being more to the performance than just the size, amount and speed.

[ChrisCross]

8 minutes ago, GER_T4IGA said:

 

yeah maxwell was pretty poor. thats why i didnt understand the titan x.

nvidia claims very much. i dont trust them at all