OTOY Forums

The question came up in another topic.

Cuda Cores are important for Octane render speed.
But it seems that different architectures (fermi, kepler, etc.) do not have the same CUDA core defintion, i.e. if I buy a Kepler Card with 1000 CUDA cores it is not the same as Fermi 1000 CUDA cores.

What is a CUDA? A virtual oder a physical thing? An performance standard (number of operations/time) or a collection of transistors that I can hold in my hand? Does it have a unit, like bandwidth and so on? All I know so far is more is better, but more of what...Kepler, Fermi, Newton, Laplace ... no idea.

Maybe someone knows? So far it could be anything...

Check it out here:
http://www.nvidia.com/object/cuda_home_new.html

I know the marketing stuff, but it does not say what an actual 'CUDA core' is, how it is defined...

A CUDA core is just a shader processor of which many (i.e. 32/192/128) sit on a streaming multiprocessor (SM/SMX/SMM) of which multiple sit on a GPU die. They don't work independently which means there are limitations how kernel instructions can be executed. These limitations depend on the architecture. For an overview over some architectures see for example:

http://www.anandtech.com/show/5699/nvid ... 0-review/2
http://www.anandtech.com/show/8526/nvid ... 0-review/2

Thanks, the second link explains it quite well.
A fast conclusion would be that Maxwell 980 is slower because of the part
'But logically the execution blocks which each warp scheduler can access have been greatly curtailed.', which could be considered as one of the limitations, in addition to the overall lower cuda core number. Just a guess, would be interesting to get it laid out in more detail.

well, I guess you don't just reduce power consumption without sacrificing on some end.