Ask HN: Build a deep learning box from scratch or just buy one? | Hacker News Reader

Ask HN: Build a deep learning box from scratch or just buy one?

10 pointsmisiti37807 comments

I'm interested in hearing from anyone here who has either built a deep learning box from scratch or purchased one? I would like to build one from scratch but if the analogous HW can be purchased already constructed for a similar amount of money, I would just buy it.

I'm basically looking for links to instructions for building your own box and/or links to a product I could purchase.

View on HN

I'm not exactly sure what a 'deep learning box' is, but if I wanted a box to stuff some GPU's on a budget I'd start with a used Dell Precision 7xxx series off of ebay for a couple of hundred dollars as a platform.

These machines have two graphics x16 PCIe slots wired as x16 and two more x16 wired as x8 powered by an 1100 watt power supply. New, a similar power supply would cost more than the what a whole system will go for including a Xeon CPU, some Ram and possibly a disk and perhaps a Windows license.

The downside is that the Precision 7xxx's are big and heavy and a bit power hungry and only come in Oakland Raiders black with silver.

The Precisions are also certified for various flavors of Linux when new, so there's unlikely to be any driver problems regarding their hardware. They're stupid easy to open up and in and around...a similar new case would also probably cost more than a complete used system. I've found them very reliable.

From there, it would just a be matter of how much money there is for GPU/graphics cards, more RAM, and SSD's and how it gets spread around based on the anticipated workload: a $50 SSD and a $50 GPU might be enough for fooling around and developing a toolchain. After that, it's just a matter of how far you can fall when sitting on your wallet.

Adrian Rosebrock of PyImageSearch explained his rationale for purchasing a pre-configured Nvidia box in [1] and talked about other build considerations in [2] (I'm not affiliated in any way; just a subscriber of his blog who thinks you may find the info there useful).

[1]: http://www.pyimagesearch.com/2016/06/06/hands-on-with-the-nv...

[2]: http://www.pyimagesearch.com/2016/06/13/considerations-when-...

To build mine I took an existing Linux server (nothing special, 16G RAM and a 3.something CPU) and squeezed a GTX 980 into its tiny frame. The speed up for basic image processing was about 31x over CPU alone.

I don't know anything about pre packaged solutions but maybe a gaming rig would be appropriate.

This is probably the most current at-home option. It's basically just building an average gaming rig, yes you can buy them prebuilt but you probably just need an i5 or i7 a good pool of RAM and a late-model GTX. There's many ways prebuilt or otherwise to accomplish that.

Here is an example of a prepackage solution: http://exxactcorp.com/index.php/solution/solu_detail/225

seems expensive

Here's one guy's approach: http://www.kdnuggets.com/2016/06/build-deep-learning-box.htm...

You can go to drive-vector.com They take order for Deep Learning Box for researchers and self-driving car developers

Reply on news.ycombinator.com