CODEWITHSUNDEEP
cntk
Ujjwal
Ujjwal

Reputation: 1859

CUDA Compute capability in CNTK

During the installation of CNTK (Version 2.0), there appears to be no option for setting the compute capability. I am curious, as to how is the compilation of CUDA modules in CNTK, optimized for graphics cards with different compute capabilities ? One can have two different sets of CUDA cards (with different compute capabilities) on two different nodes, and hence this question naturally comes to mind.

Upvotes: 0

Views: 128

Answers (2)

Emad Barsoum
Emad Barsoum

Reputation: 476

For our pure cuda code, we don't take advantage of the various compute capability as mentioned by Nikos. However, we also depend on cuDNN, for all operations that is supported by cuDNN, and cuDNN take advantage of the various computation abilities.

Upvotes: 1

Nikos Karampatziakis
Nikos Karampatziakis

Reputation: 2050

This line suggests that CNTK is not taking advantage of Pascal cards' compute capabilities. I think the reason for this is we don't have enough of these cards in our continuous integration system.

Upvotes: 1

Related Questions