ruffsl Wed, Jul 22, 2015 at 4:26 PM
To: kaixhin.com, couchbase.com
Hello Traun and Kai,
I’ve started a thread discussing the creation of an official Docker Hub repo for CUDA. I’ve witnessed you both are very active online in using Docker and CUDA (and thanks for paving the way with documentation), so I wanted to invite you join the discussion:
Official Docker image for CUDA? - Announcements - NVIDIA Developer Forums
Thanks,
ruffsl
Traun Leyden <couchbase.com> Thu, Jul 23, 2015 at 5:02 PM
To: ruffsl
Cc: kaixhin.com
Hi ruffsl,
Thanks for sending that. Once that official DockerHub repo is out, I will switch elastic-thought over to use it.
Also, I updated my DockerHub README to have a link to your nvidia devtalk forum post.
ruffsl Thu, Jul 23, 2015 at 5:25 PM
To: Traun Leyden <couchbase.com>
Cc: Kai Arulkumaran <kaixhin.com>
Hi Traun,
Wow, hadn’t noticed your elastic-thought project yet, going to have to read into that soon.
Mind if you chime in on the form? Looks like the Nvidia board is a bit lurker community. Other post have nearly ~13,000 views, but only 13 replies, and at that ratio, I don’t think I’ll be getting 1,000 views of traffic for a reply anytime soon.
Perhaps you’d like to recommend what tags should be published, cuda:6.5, cuda:7.0? Should cuDNN be included? Have a candidate example for Dockerfile setting up CUDA?
I’ve poked an academic friend I know at Nvidia to get some official involvement, (I’d like to make sure the official repo would be well maintained as it requires a bit of commitment), but I haven’t gotten any word back on my end. You wouldn’t know any contacts there, would you?
Thanks,
ruffsl
Kaixhin <kaixhin.com> Fri, Jul 24, 2015 at 6:01 AM
To: ruffsl
Cc: Traun Leyden <couchbase.com>
Hi ruffsl,
Unfortunately I’m abroad on holiday at the moment, so I’ll have a better look once I get back.
I agree that cuda:6.5 etc would be a good tagging system, though you may need an extra distinction for those who want the samples installed as well. One restriction is that any machines need the appropriate driver version as the one that would be installed in the image, and I’m not sure NVIDIA have a way around this. So even though an official image would be great, I wonder if NVIDIA will consider it worth investing in, given the extra support needed. I’ve asked my contact at NVIDIA about this before, but he doesn’t really have anything to do with this area.
cuDNN is under some kind of licence which means that they are unlikely to include it in an image (for now); it’s also far more specific than any of their included libraries.
Best,
Kai
ruffsl Fri, Jul 24, 2015 at 3:17 PM
To: Kaixhin <kaixhin.com>
Cc: Traun Leyden <couchbase.com>
Hey Kai,
In regard to your comments,
I’ve added those topics to a new reply in the current devtalk thread.
Thanks,
ruffsl