CODEWITHSUNDEEP
windowsgpunvidiacaffetesla
Mona Jalal
Mona Jalal

Reputation: 38165

Caffe on Windows common.cpp:166] Check failed: error == cudaSuccess

When I use two different GPUs for Caffe win7, it does not work. OS system: win7 SP1 GPUs: Quadro K620, Tesla K40 Caffe: Caffe for windows The problem is that every time I update to the latest Driver 369.26 for K620, then Tesla K40 will not work. And then when I update to the latest Driver 341.96 for Tesla K40, then K620 will not work. So it does not work for caffe when I use dual GPUs and the error is: 

C:\Softwares\caffe-windows>.\Build\x64\Release\caffe.exe train --gpu=0,1 --solve
r=./examples/mnist/lenet_solver.prototxt
I0919 22:42:21.715857  5568 caffe.cpp:186] Using GPUs 0, 1
I0919 22:42:21.725857  5568 caffe.cpp:191] GPU 0: @■0
I0919 22:42:21.725857  5568 caffe.cpp:191] GPU 1: @■0
F0919 22:42:21.725857  5568 common.cpp:160] Check failed: error == cudaSuccess (
35 vs. 0)  CUDA driver version is insufficient for CUDA runtime version
*** Check failure stack trace: ***

And if I just use GPU Tesla K40(use Quadro only, it works), the error is:

C:\Softwares\caffe-windows>.\Build\x64\Release\caffe.exe train --gpu 1 --solver=
./examples/mnist/lenet_solver.prototxt
I0919 22:20:00.002020  4168 caffe.cpp:186] Using GPUs 1
I0919 22:20:00.095620  4168 caffe.cpp:191] GPU 1: @■/
F0919 22:20:00.095620  4168 common.cpp:166] Check failed: error == cudaSuccess (
10 vs. 0)  invalid device ordinal
*** Check failure stack trace: **

Upvotes: 0

Views: 674

Answers (1)

Xiong Yunyang
Xiong Yunyang

Reputation: 51

I think that the problem is every time on Win7, nvidia update will integrate two drivers to one. Then if one driver is higher for one and caffe asks two drivers to the most updated. Thus, it will not work.

Upvotes: 2

Related Questions