Tesla V100 VS. P100 (deep learning speed testing)

[dhenzjhen]

This is only a speed training testing without any accuracy or tuning involved. I only want
to test and compare the V100s and P100s in terms of crunching speed. In this testing, I used
1281167 training images and 50000 validation images (ILSVRC2012) and NV-caffe for deep
learning framework. On the networks, I used AlexNet, GoogleNET and CaffeNET with different
batch sizes and 300,000 iterations runtime on all the tests.

System: Supermicro SYS-4028GR-TXRT
Motherboard: X10DG0-T
CPU: E5 2699V4 x 2
MEM: 32GB Micron x 12
BIOS: 5/25/17
GPU: Nvidia Tesla V100 SXM2 x 8 | P100 SXM2 x 8
OS: Ubuntu 16.04 x64
Driver: 384.81
CUDA: version 9

-dhenzjhen@SMCI

 

[William]

WOW man, pretty sweet !

 

[dhenzjhen]

WOW man, pretty sweet !

Thank u sir!

 

[hifijames]

That's some serious horsepower you got there! Wish I could have that kind of toys to play with
I am a total noob here, so please forgive my ignorance. How involved will it be to switch one framework to another, say from Caffe to Tensorflow?

 

[gigatexal]

yeah yeah but can it run crysis? jk, that's pretty sweet.

 

[LukeP]

Looks terrible. V100 should be 10 times faster

 

[LukeP]

any followup why V100 with its 110Tflops is so slow? was that benchmark not tuned for V100?

 

[dhenzjhen]

any followup why V100 with its 110Tflops is so slow? was that benchmark not tuned for V100?

Hi Luke,

Sorry I don't have access to GPU servers anymore.