compare px2 and gtx950，the result shows px2 is much weaker than gtx950, what is the problem I have m...

Autonomous Machines DRIVE PX2 General
#1

Hi, I have compared the performance between px2 and my computer with gtx950.
I did like bellow.
deeplearning platform: caffe
dataset: mnist
net: lenet

yes，I just do a basic example in caffe.

I test with command “./build/tools/caffe time --model=examples/mnist/lenet.prototxt --gpu 0”

the result like below:
gtx950 px2
data forward: 0.00146304 0.00739456ms
data backward: 0.00144768ms 0.00909312ms
conv1 forward: 0.687928ms 3.34672ms
conv1 backward: 0.421037ms 3.21993ms
…
average forward pass: 1.53327ms 7.44161ms
average backward pass:1.02709ms 6.92709ms
total time : 130.293ms 731.213ms

form the result, I have to get the conclusion that px2 is much weaker than gtx950.
so, am I right? anyone could help me? I will be so appreciated!!

#2

Dear mjjdick,

Can I know the reason why you are trying to compare the perf between GTX950 and DPX2?
Because they have different architectures, memory, CUDA core, clocks, etc. Thanks.

#3

Because we have a deeplearning project, the performance and cost just be OK in gtx950. Now we need to deploy in px2 for auto drive, and we found it cost too much. so we want to know if we have a bad use for px2.

#4

Dear mjjdick,

DPX2 is for Autonomous Driving development platform not only for deployment device.
DPX2 support various automotive interfaces like GMSL camera, CAN, LIN and dual system for redundancy and Driveworks for autonomous SW development.
So I don’t think it’s enough to simply compare DPX2 with network performance only. Thanks.

#5

em… so, is there a solution for deeplearning project used in px2？ as our net cost more than 500ms per frame , even we use tensorrt and half precision , I think the cost is still too much…