Building a Deep Learning (Dream) Machine | Hacker News Reader

Building a Deep Learning (Dream) Machine

graphific.github.io

95 pointsiamjeff21 comments

Part 2: https://graphific.github.io/posts/running-a-deep-learning-dr...

Wrong stance on SSD: always get an SSD where to put the temporary data, i.e. the form that your dataset takes to be fed into the DL backend. For instance, working with Caffe, your image dataset is turned into an LMDB or HDF5 database, while you can keep your unprocessed dataset on a regular HDD. The trick is to always prevent your GPU(s) from starving.

Good build. If you have cash to burn and don't want to do a build, you can buy a prebuilt one with 4x new titan X in the UK - https://www.scan.co.uk/3xs/configurator/nvidia-deep-learning...

My projects involve a blend of fast CPU and GPU, so I decided to build a fast system recently with 1 GeForce GTX 1080, a fast Intel Core i7, and 128 GB of RAM.

I'll install another GTX 1080 soon. I found that the new Titans are hard to get and the performance/price ratio seems favorable on the 1080's.

I got Ubuntu 16.04 to work with the latest 1080 graphics drivers and CUDA 8 RC drivers -- no small feat. Reading all the documentation helps, and even then, it took some finagling.

See https://github.com/beniz/deepdetect/issues/126#issuecomment and other links online. The trick is to install CUDA 8 without the attached driver, then install the new NVidia driver. Once you got that, install on 16.04 is straightforward.

Yep. Also, this is a nice write-up for TensorFlow with CUDA 8.0 RC and cuDNN 5.1.5 with Ubuntu 16.04:

https://alliseesolutions.wordpress.com/2016/09/08/install-gp...

if CPU isn't that important, then I suggest getting a second had hp Z800, with the big PSU

lots of space for GPUs lots of space for ram and disks.

Are there any hosting solutions that offer GPUs yet?

http://www.nvidia.com/object/gpu-cloud-computing-services.ht...

Microsoft Azure's K80 based N-series are the best that I've seen so far for by the hour instances. They're currently in preview availability[1].

Compared to the best AWS GPU, the Azure K80s are (a) cheaper and (b) many times more performant for the purposes of deep learning.

Hopefully this lights a fire under AWS to become more competitive with their GPU range.

[1]: https://azure.microsoft.com/en-us/blog/azure-n-series-previe...

I'm still on the waiting list for the new N-series GPU servers. Does anyone have any idea about when it will be made available more broadly?

what was the total price of this box - did i miss the price in the article ?

I spent almost $2400 on my box. Finished it just a few weeks ago. I used parts of this article for inspiration. I only have 2 x NVIDIA GTX TITAN-X 12GB (bought used), not 3. I also used a 1600w power supply.

How does the Titan compare to the new 1080s for machine learning?

Slower but more memory (8GB vs. 12GB). Depending on the problem it could definitely matter.

There is also a new pascal based titan.

for my needs these cards are working just fine. I could post a picture of the rig if anyone wants to see it.

Convnet benchmarks say that a 1080 is only like 5% faster. 12GB of VRAM seemed more important. I just spent yesterday evening sticking a cooling rig onto a Titan X I bought on eBay. Great value atm if you're willing to get your hands dirty. Will admit that it was a struggle though.

Recently I built a similar machine with a single GTX 1080 for $2500 in Norway.

It's about $6500 ( http://pcpartpicker.com/list/y7MrvV )

Those GPUs are 1200 USD apiece, the rest seems like it would be about another 1200, so I'd say 4 to 5 grand realistically. The nice thing is that since the majority of the cost is focused on the GPUs, you can build the rest of the box and add GPUs later as needed.

That's pretty much right on. I just built a box with one Titan X and the total cost was about $2400.

Reply on news.ycombinator.com