806
75%

Ran 19 Dec 2015 03:57PM UTC

Jobs 1

Files 106

Run time 3s

Badge

Embed ▾

pending completion

Build # 806

Build Type

Pull #480

travis-ci

Committed by

gheinrich

Commit Message

Torch Multi-GPU support

This commit enables data parallelism, i.e. train
batches are evenly spread across selected GPUs.

Strong scaling is applied by default, i.e. batch
size is unchanged and each GPU receives its share
during training.

For better performance, installation of the NCCL
module is recommended.

On a 2xTitanX machine (with NCCL):
- GoogleNet: 51m4s (1 GPU) -> 34m3s (2 GPUs)
- Alexnet: 12m31s (1 GPU) -> 10m38s (2 GPUs)
- LeNet*: 53s (1GPU) -> 1m30s (2 GPUs)

* the overhead of inter-GPU communication seems to
outweigh the compute gain on LeNet

Pull Request Pull Request #480: Torch Multi-GPU support

Run Details

7886 of 9762 relevant lines covered (80.78%)

0.81 hits per line

Uncovered Existing Lines

Lines	Coverage	∆	File
3	92.21	0.0%	digits/frameworks/torch_framework.py
6	95.16	0.38%	digits/model/images/classification/test_views.py
48	80.83	0.29%	digits/model/tasks/torch_train.py

Jobs

ID	Job ID	Ran	Files	Coverage
1	806.1	19 Dec 2015 03:57PM UTC	106	80.78	Travis Job 806.1

Source Files on build 806

Coverage	∆	File	Lines	Relevant	Covered	Missed	Hits/Line

NVIDIA / DIGITS / 806
75%

README BADGES
x

Markdown

Textile

RDoc

HTML

Rst

Uncovered Existing Lines

Jobs

Source Files on build 806

NVIDIA / DIGITS / 806 75%

README BADGES x

Markdown

Textile

RDoc

HTML

Rst

Uncovered Existing Lines

Jobs

Source Files on build 806

NVIDIA / DIGITS / 806
75%

README BADGES
x