Deep learning with 4028GR-TXR (AlexNet)

[dhenzjhen]

Deep learning for fun and new *opportunities

Supermicro 4028GR-TXR system
Intel E5 2698 V4 x 2
Samsung 16GB 2400 x 24
Nvidia P100-SXM2 GPU x 8
BIOS 3/22/17
IPMI 3.52
Software stack: Centos 7.2 x64, Nvidia driver 375.20, Cuda_8.0.44, Caffe(AlexNet/GoogleNet), Nccl
Glog, Lmdb, Hdf5, Cudnn-8.0, Gflags, Python and Leveldb

All results are based on Alex Krishesky's published papers.

References:
NVIDIA Caffe · NVIDIA/nvidia-docker Wiki · GitHub
How to Download and Install Caffe on NVIDIA GPUs
Model Zoo · BVLC/caffe Wiki · GitHub
Caffe | Model Zoo
Caffe | ImageNet tutorial

 

[William]

Nice DJ !

 

[dhenzjhen]

Thanks Bill! I need to try more on 2x and 1x config to get close to at least 77-78% accuracy.

 

[MiniKnight]

We need to docker this. I want to try on my new 1080 Ti. Maybe @Patrick can run on his new beast? Deeplearning010 - 8x GTX 1080 Ti Build (Deep learning and crypto mining)

 

[Patrick]

@MiniKnight agreed.

@dhenzjhen any chance I can pick your brain on setting this up next week?

 

[dhenzjhen]

Yessir no prob, but if you want to start asap there's a Nvidia caffe link on the reference you can follow.

PS. I know KJ by the way


Sent from my SM-N900T using Tapatalk

 

[MiniKnight]

MOAR Benches! @Patrick you should do a youtube vid of one of those systems.

 

[dhenzjhen]

Moar benchies coming soon for sure

 

[Patrick]

Just woke up. 4:30AM Taipei time.

@dhenzjhen - next week item. I also edited KJ's name to initials.

 

[dhenzjhen]

I tried training images with digits and it's fun but of course the backend still using caffe
Index of /~chmod755/sth

 

[Patrick]

I tried training images with digits and it's fun but of course the backend still using caffe
Index of /~chmod755/sth

Digits is easy but not used as widely as NVIDIA would lead some to believe.