Search code examples
cntk

Cntk can not detect my gpu by cntk.all_devices()

Cntk only detect 1 device(my cpu) by calling cntk.all_devices(). However I do have a gpu on my computer. By running the tutorial supported by cntk, I could get some info:

-------------------------------------------------------------------
-------------------------------------------------------------------
GPU info:

                Device[0]: cores = 48; computeCapability = 2.1; type = "NVS 310"; memory = 512 MB
-------------------------------------------------------------------

##############################################################################
#                                                                            #
# Train command (train action)                                               #
#                                                                            #
##############################################################################


Model has 9 nodes. Using CPU.

As an consequence, I can not use my gpu as by calling set_default_device(gpu(0)). How could I solve this problem?

Solution

  • The minimum GPU compute capability for CNTK is 3.0. (Edit: The fact that you can run the tutorial using cntk.exe indicates a bug somewhere in the v1 executable.). When you run the tutorial with cntk.exe, it prints out the GPU info, but still ends up using CPU: Model has 9 nodes. Using CPU.

    The only way to solve this problem is to change the value of the constant MininumCCMajorForGpu in BestGpu.cpp and recompile.