[
https://issues.apache.org/jira/browse/SINGA-308?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15963926#comment-15963926
]
wangwei commented on SINGA-308:
-------------------------------
Great!
The gpu worker and cpu worker sync every iteration.The gpu is very fast and
spends most time waiting for the cpu. Therefore the speed is similar to CPU
training.
You can update the updater to do asynchronous training, which run gpu and cpu
asynchronously to avoid the synchronization overhead.
refer to this file:
https://github.com/apache/incubator-singa/blob/master/src/model/updater/local_updater.cc#L55
> CPU-GPU parallelism
> --------------------
>
> Key: SINGA-308
> URL: https://issues.apache.org/jira/browse/SINGA-308
> Project: Singa
> Issue Type: Test
> Components: Core, PySINGA
> Environment: Ubuntu 16.04
> CPU-GPU of the same machine
> Reporter: Muhammad Hamdan
> Labels: test
>
> Is it possible to parallelize the alexnet model for the cifar10 example on a
> CPU and GPU instead of 2-GPUs ? Assuming asynchronous communication between
> the two components
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
