This is my first deep learning project. Besides I gained a lot of experience, I have found relatively high cost associated with GPU computing. My GPU (GTX 980) started working since Feb, which would cost me $120 ~ 180 in electrical billing. By normalization by 6 weeks, $3~4/per day.
", "messages": [ { "id": "66518", "postDate": "03/16/2015 21:30:37", "content": "This is my first deep learning project. Besides I gained a lot of experience, I have found relatively high cost associated with GPU computing. My GPU (GTX 980) started working since Feb, which would cost me $120 ~ 180 in electrical billing. By normalization by 6 weeks, $3~4/per day.
", "rawMarkdown": "", "votes": null }, { "id": "66525", "postDate": "03/16/2015 21:41:14", "content": "I have GTX 970 and for this competition it ran almost 24/7. It cost me around $2-3/per day.
\nAmazon GPU spot instances will be cheaper ($0.0649 per Hour, see: http://aws.amazon.com/ec2/purchasing-options/spot-instances/).
", "rawMarkdown": "", "votes": null }, { "id": "66530", "postDate": "03/16/2015 21:45:22", "content": ">Amazon GPU spot instances will be cheaper
Haven`t tried for this competition, but from previous experience Amazon GPUs are slower than GTX770/780.
This is also my first deep learning project, I am running all the caffe code on my laptop with a GT755M GPU, the largest model I run took two days to finish, because it's a laptop, I had to stop the code several times during training to cool down the gpu. Right now I just hope it will not shorten the life of my laptop too much.
\nI wish I had a titan z ...
", "rawMarkdown": "", "votes": null }, { "id": "66562", "postDate": "03/16/2015 23:16:26", "content": "Also first deep learning project. I used ec2 spot instances and they were around $.065 / hr. They use the GRID 520K. Aren't those some of NVIDIAs faster graphics cards? I wasn't sure.
", "rawMarkdown": "", "votes": null }, { "id": "66583", "postDate": "03/17/2015 00:29:14", "content": "At times I run 15 machines on ec2 in parallel trying different options, this is something you cant reproduce if you are running on private hardware
", "rawMarkdown": "", "votes": null }, { "id": "66599", "postDate": "03/17/2015 01:15:53", "content": "EC2 spot instances worked reasonably well and cheaply, although I got burned in the last week when I spun up several instances to try some last minutes tweaks and the spot price suddenly spiked 5x a couple days later, terminating all of them.
", "rawMarkdown": "", "votes": null }, { "id": "66727", "postDate": "03/17/2015 15:51:15", "content": "[quote=Matt Krump;66562]
\nAlso first deep learning project. I used ec2 spot instances and they were around $.065 / hr. They use the GRID 520K. Aren't those some of NVIDIAs faster graphics cards? I wasn't sure.
\n[/quote]
\n\nIt can be figured out. For Caffe compiled with CUDNN its MNIST example takes ~50 seconds on AWS.
", "rawMarkdown": "", "votes": null }, { "id": "68487", "postDate": "03/26/2015 21:33:07", "content": "I tried out AWS too and liked it, but by that time there was only one day left before the end of the competition! Anyway, here's my timing experience.
\nFor cxxnet, as originally posted:
\nGeforce 780: 5 minutes (according to the README --- I didn't have this card to try)
\nGRID 520K: 10 minutes (my experience on AWS)
\nGT 550M: ~65 minutes (on a laptop I borrowed --- gpu from mid-2011)
\ni7 cpu: 7 hours :) (same laptop)
\nBTW, those GPU times are roughly in line with the scores at http://www.videocardbenchmark.net/high_end_gpus.html
\nSo the AWS cards are about half as fast as the high end cards, but they still seem like a decent deal to me. At least for beginners to learn about CNNs. :)
", "rawMarkdown": "", "votes": null }, { "id": "68507", "postDate": "03/27/2015 01:08:41", "content": "For me, given the price electrical disruptions and accessibility I think AWS is the winner. It will run slowly but for me it is reliable than having GPU at home (of course you have to put reasonable spot price so that it won't get terminated, happened to me only once in the last couple of months)
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 66525, "author_name": "dmitrylarko", "author_url": "", "post_date": "03/16/2015 21:41:14", "content": "I have GTX 970 and for this competition it ran almost 24/7. It cost me around $2-3/per day.
\nAmazon GPU spot instances will be cheaper ($0.0649 per Hour, see: http://aws.amazon.com/ec2/purchasing-options/spot-instances/).
", "votes": null, "replies": [] }, { "id": 66530, "author_name": "oldufo", "author_url": "", "post_date": "03/16/2015 21:45:22", "content": ">Amazon GPU spot instances will be cheaper
Haven`t tried for this competition, but from previous experience Amazon GPUs are slower than GTX770/780.
This is also my first deep learning project, I am running all the caffe code on my laptop with a GT755M GPU, the largest model I run took two days to finish, because it's a laptop, I had to stop the code several times during training to cool down the gpu. Right now I just hope it will not shorten the life of my laptop too much.
\nI wish I had a titan z ...
", "votes": null, "replies": [] }, { "id": 66562, "author_name": "matthewkrump", "author_url": "", "post_date": "03/16/2015 23:16:26", "content": "Also first deep learning project. I used ec2 spot instances and they were around $.065 / hr. They use the GRID 520K. Aren't those some of NVIDIAs faster graphics cards? I wasn't sure.
", "votes": null, "replies": [] }, { "id": 66583, "author_name": "udibr1", "author_url": "", "post_date": "03/17/2015 00:29:14", "content": "At times I run 15 machines on ec2 in parallel trying different options, this is something you cant reproduce if you are running on private hardware
", "votes": null, "replies": [] }, { "id": 66599, "author_name": "telser", "author_url": "", "post_date": "03/17/2015 01:15:53", "content": "EC2 spot instances worked reasonably well and cheaply, although I got burned in the last week when I spun up several instances to try some last minutes tweaks and the spot price suddenly spiked 5x a couple days later, terminating all of them.
", "votes": null, "replies": [] }, { "id": 66727, "author_name": "rakhlin", "author_url": "", "post_date": "03/17/2015 15:51:15", "content": "[quote=Matt Krump;66562]
\nAlso first deep learning project. I used ec2 spot instances and they were around $.065 / hr. They use the GRID 520K. Aren't those some of NVIDIAs faster graphics cards? I wasn't sure.
\n[/quote]
\n\nIt can be figured out. For Caffe compiled with CUDNN its MNIST example takes ~50 seconds on AWS.
", "votes": null, "replies": [] }, { "id": 68487, "author_name": "selimte", "author_url": "", "post_date": "03/26/2015 21:33:07", "content": "I tried out AWS too and liked it, but by that time there was only one day left before the end of the competition! Anyway, here's my timing experience.
\nFor cxxnet, as originally posted:
\nGeforce 780: 5 minutes (according to the README --- I didn't have this card to try)
\nGRID 520K: 10 minutes (my experience on AWS)
\nGT 550M: ~65 minutes (on a laptop I borrowed --- gpu from mid-2011)
\ni7 cpu: 7 hours :) (same laptop)
\nBTW, those GPU times are roughly in line with the scores at http://www.videocardbenchmark.net/high_end_gpus.html
\nSo the AWS cards are about half as fast as the high end cards, but they still seem like a decent deal to me. At least for beginners to learn about CNNs. :)
", "votes": null, "replies": [] }, { "id": 68507, "author_name": "akilaw", "author_url": "", "post_date": "03/27/2015 01:08:41", "content": "For me, given the price electrical disruptions and accessibility I think AWS is the winner. It will run slowly but for me it is reliable than having GPU at home (of course you have to put reasonable spot price so that it won't get terminated, happened to me only once in the last couple of months)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "66518": "", "66525": "", "66530": "", "66560": "", "66562": "", "66583": "", "66599": "", "66727": "", "68487": "", "68507": "" }, "source": "meta" }