Deep learning box for $1700 | Hacker News Reader

Deep learning box for $1700

blog.slavv.com

73 pointspplonski8634 comments

FYI the OP also created a really nice way to create persistent spot instances on AWS - see approach 2 here: http://wiki.fast.ai/index.php/AWS_Spot_instances

If the boxed fans have not improved significantly in the last years then a good heatsink/fan would have been a very good investment. At least if one sits in the same room where the box is running.

Wrong CPU. Only 16 PCI lanes.

Would you mind to elaborate?

It's not enough bandwidth for multiple GPUs.

PCIe lane width isn't normally a significant bottleneck for GPU performance, for games at least. Going from 16x to 8x loses you 1% or so, depending on the hardware. I don't know the bus traffic profile of software being run for deep learning, but it may not be noticeable.

I'd also pick a CPU with more lanes though.

It's fine for a single GPU. For multiple GPUs in a deep learning box you want more.

What cpu would you recommend? Would it be better to go with a 2x cpu mobo?

Definitely not since only 1 socket is usually directly attached to the PCIe bus. The best CPU for the money at the moment is most likely an i7 5830K with 40 PCIe lanes. Or you could wait (probably the best idea) for AMD to release it's "threadripper" CPUs tomorrow with 44 lanes and for Intel to release it's new i7/i9 lineup in 2 weeks, also with 44 lanes.

Cool, thanks for the info.

Agreed. This would bifurcate down to 8/4/4 if you wanted to add an NVME SSD or 10Gbit NIC

Hi, I'm the author. My thought was that running 2 GPUs in a 2x8 config wasn't going to be a bottleneck. Or is it? Apparently my understanding of PCIe is totally off. What would be a suitable CPU (with 40 lanes?) for dual GPU in this price range ($200-$300)?

Thank you for the comments, eventually I can stop spreading my ignorance around :)

I went for a slightly cheaper option ($950), with a slower GPU, less disk, but more RAM and better PSU and case:

- Used Dell T7500 (Xeon X5675, 48GB RAM): US$350 including delivery from ebay: http://www.ebay.com/itm/Dell-T7500-Home-PC-Xeon-6-Core-3-06G...

- Nvidia GTX1080 8GB (US$480)

- SSD ($100)

- USB 3.0 PCIE card ($20)

Considering the ridiculous amounts of time spent on training, how do those of you deal with 'work' ? Work remotely ? Run multiple experiments at the same time ?

I'm having an increasingly difficult time explaining to my client why it's taking so much longer to debug the code as opposed to getting the implementation off the ground.

This is 100% the process I've gone through to build a DL box. Anaconda saved me in he fast.course.ai lessons.

How did you like those courses?

It's OK but definitely think we need a bigger picture / organically growing course material.

Just to clarify - the actual url is course.fast.ai :)

haha, sorry pre-coffee! :)

If the OP reads this, this fragment makes no sense:

> Even thought the GPU is the MVP in deep learning

Also: if you ever do another build: do not take the parts out of their protective packaging until you need them. A nice little bolt of static from your fingertips to the motherboard, memory sticks or GPU terminals could ruin the part. Those bags and the foam are conductive and help to disperse static energy so better to use them until the last moment.

MVP as in Most Valuable Player.

Most people would read that abbreviation as 'Minimum Viable Product' in the HN context.

Not me, I totally understood what he meant due to the context. The reality is that "Minimum Viable Product" was hijacked from the MVP of sports.

I'm not into sports (unless I'm doing something myself :) ), so I never heard it used like that before.

Its a pretty common Americanism that is disconnected from sports in my experience, I suspect the average American would recognize it as "something/someone that is contributing a lot", and only a tiny percentage would even recognize "minimum viable product" spelled out.

Or outside the US - "MVP" seems to come from US sports terminology which are still relatively niche in the rest of the world (compared to Football anyway, and I reckon there's a large amount of people who don't even care about sports)

Heh, there's a painting company that parks their van along Kezar Drive in SF with the name/logo "MVP", which they intend to be read as Most Valuable Painter. It always makes me think of the scene in an old episode of Mr Bean where he paints a room by putting a firecracker in a paint can.

Thanks for the tip! You're the real MVP, @jacquesm ;)

I have thought about the same thing when playing with deep learning at home, until I put electricity bills into the equation. A beast like what the author has built, with full power, will cost $0.1/hour in electricity bill alone at where I live ... which almost equal to the cost of a p2.xlarge spot instance on Amazon.

So if someone else is paying for the electricity, build your own gig, otherwise, Amazon is pretty hard to beat.

Same goes for Bitcoin mining.

The spot pricing is pretty crappy to compare to since you only get to use spare capacity. If you have any project where availability matters then you should compare to the on demand price of $0.9/hour.

If availability doesn't matter, it means you wouldn't run the box full time anyway, in which case your power calculation is too high.

It's a processing task. You run it for a fixed CPU time. Wall time may or may not matter depending on the application.

You almost never take availability into account, but taking spot prices will increase wall time.

Except his build is 4x faster: https://medium.com/initialized-capital/benchmarking-tensorfl...

You can get a 12-core Xeon (evaluation sample) v3 for ~$200.

Reply on news.ycombinator.com