Building Machine Learning Machines

0

Well if you somehow can’t afford the $128K Pascal server that nVidia just announced (I can’t imagine why), what should someone do if they want to do some machine learning?
Well, unfortunately, there are no cheap good alternatives right now as we are in the middle of the Maxwell to Pascal transition with nVidia and they are only shipping out super high end systems, so the main thing to do is to build a chassis and know that you will likely get rid of the current graphics cards. In other words, don’t future proof the cards, get just what you need.
Here is the recommendations on PC Parts Picker.

Graphics Cards

This is a $5-6K machine, but most of the cost is in the graphics cards. I’d say get one or two cards to start and see how your load is. The most important variable is whether you need 6GB VRAM or 12GB VRAM. The GTX 980Ti is the bargain unit here at half the price of the Titan X, but if you are doing deep neural networks, you may have to use the Titan X.
In the actual build, While you can do heterogenous deployments, I think most folks would just get either one or the other and double it up if they need more but the choices are:

  • ASUS Strix 980 Ti GTX. These are quiet cards and run well. Hopefully your load will work in a 6GB VRAM if so go for these.
  • The Titan X are hard to come by the ASUS Titan X is slightly factory overclocked, but should reach 1.2GHz if you need it. The prices are roughly $1K each.

While the graphics cards will change, the system is designed so that you can use the next generation easily enough.

Motherboard

This is the part that lasts the longest and the ASRock X99 OC Formula has all the latest features including USB 3.1, NVMe and two M.2 slots

Power Supply

The machine has enough headroom thanks to a gigantic power supply to run 4 of these cards plus the 140 watt cpu. Although it doesn’t have the normal 20% extra capacity that you want

Processor

The Xeon E5-1620 V3 overclocks and can get to 4GHz easily so it is a bit of a ways from the 4.4-4.5GHz I can get on Skylake or even 4.8 GHz if you go oil cooled on the older Sandy Bridge, but still pretty good and it gives you 40 lanes of PCI Express which you can’t get with Skylake yet.
An alternative build is to use the Core i7 Haswell-E. This is nearly an identical build but you can’t use ECC memory which matters for reliability for $10 more. Not a bad deal.
It doesn’t look like the Skylake Xeon parts will be that much faster as they are really power optimized and in a system where 80% of the power is drawn by the GPUs that really doesn’t matter as much.
The one thing that you are giving up

Memory

Well life was easy with desktop systems. Just tell me how much memory you need. The maximum is 64GB though and for big jobs you pay for more. Plus you want ECC so that you get protection from bit flips.
There are no less than three different kinds of ECC ram.

  • Unbuffered. UDIMM. These are very fast but have a maximum memory limit
  • Registered. RDIMM. These are slower by about 12% but support double the memory of UDIMMs.
  • Low load registered. LRDIMM are slower, but have double the density of RDIMM but this is not compatible with the X99.

The main reason for this is that LRDIMM is higher density but is slower and more expensive. DDR3 maxed out at 8GB in UDIMM per memory module, but DDR4 are double that density at 16GB/slot but this is very pricey. The most economical for a fully populated system would have 128GB across 8 slots with 16GB/slot in DDR4.
The ASRock OC Formula/3.1 has 8 slots for DDR4 and supports x8 with a maximum memory of 128GB, so each slot is limited to 16GB memory (16Gb chips in x8 configuration).

SSDs

This is an all SSD build (I’ve been all SSD for the last two years). This is because it is quiet and reliable. Mechanical drives definitely seem like the first thing that fail.
Make sure you update the X99 firmware so that you get full NVMe support. But this thing has two m.2 slots. One that is NVMe 4x and the other 2x PCI Express (if you have an old Plextor you can get over SATA speeds with a full 2 lane implementation). Johnnylucky.com gives you a list by interface so the Plextor M6e is $160 so only $20 cheaper than the very fast Samsung 950 Pro. I’d recommend just getting the 950 Pro and using it in the slot. You only get half the performance, but when you want to move on you get a super capable system.
For the SATA SSD storage, the SanDisk Extreme Pro remains super fast, although if you need lots of bulk storage the 1TB or if you want a little cheaper you can get $200 1TB storage as well.

Related Posts

© All Right Reserved