the speed of forward, caffe with cudnn slower than without cudnn?
BigPuns opened this issue · comments
test model: Alexnet, Vgg16, Resnet50, Restnet101,
CUDA version: cuda-9.0,
CUDNN version: cudnn7.6
Operating system: v100, 1080ti
use caffe time to test speed
./build/tools/caffe time --model alexnet.prototxt --weights alexnet.caffemodel --gpu 1
I use two modes: with cudnn and without cudnn;
when batch = 1:
with cudnn is slower than without cudnn, for example,
resnet101:
forward use time:
with cudnn : 34ms
without cudnn: 18ms
when batch = 10:
with cudnn is quicker than with cudnn, for example
resnet101:
forward us time:
with cudnn: 60ms
without cudnn: 48ms
so my question is why batch=1, with cudnn is slower? how do cudnn for conv speed up?