Caffe performance on Jetlson TK1

[Pankaj Randhe]

Hi...

I seek the little guidance on the problem I am currently facing with performance of Caffe Deep CNN for image classification on Jetson TK1.

My Caffe model is taking approx. 4 secs on ARM cpu and approx 7 secs on Jetson's GPU for forward pass during prediction phase for the batch size of 60K images with dimensions 20*5*5. Here GPU is taking more time than CPU for forward pass. However, I get about 1.2X speedup with the same code on GeForce desktop GPU with 48 cores.

What could be the reason behind this strange behaviour of Jetson TK1?

[Ihsan Ullah Khan]

Hi Pankaj,

Did you solve the problem?

I am facing the same problem on jetson TX1.

Kindly give me some solution if you solved the problem.

Thanks in advance.

Regards

Ihsan Ullah

[Davide Behr]

Hello Ihsan, 

I too am interested in the performance of Caffe on TX1, do you have any insights? 

Thank you, 

Davide 

[Ihsan Ullah Khan]

Hi Davide;

I still trying to solve that problem!

When I test the classification without using Qt Interface its performance is about 500 ms aproxx.

Now a days I am trying to deploy the Faster RCNN on JTX1.

Can you help me to solve this problem.?