My old work computer is starting to show its age, and so for about a year now I’ve been almost exclusively working on one of the computers in the lab cluster. The mini-cluster, which I built out of robust core i7 920 machines to analyze array data, has been a great success, and now it’s basically full. So I ordered a custom-built number crunching machine, paid for by my very awesome advisor, which I just received.
Here it is on the right, next to a regular-sized tower for comparison. Many thanks to Theo Zanos who took this and the other pictures. The computer is huge, and kind of looks like the monolith in 2001. It was about 2500$ (Canadian, taxes included), which is cheap relative to the machine’s specs. What follows is a hardware list and some of the considerations that went into building the machine, which I hope will be helpful to fellow scientists building their own.
What to buy and where
I like Macs, but for number-crunching machines they are simply unaffordable. A quick quote from the Mac website shows that a computer with specs comparable to the PC I built is more than 5 grand! Buying a computer pre-made from Dell or some other company carries a significant premium as well. It’s well worth it to have it custom-built. I recommend going with a small local shop as service is often better than in big-box stores and more convenient than with online shopping.
CPU, Memory and Mobo
For the CPU, I looked at the ratings off the PassMark website. The fastest CPUs are in the Intel Xeon and i7 lines. Xeons, which are intended at the server market, are very expensive. The i7s are much more affordable. I’ve had great experiences with the i7 920 in the past, a 4-core CPU clocked at 2.66 GHz. 2.5 years after its release, it’s still more than half the speed of the fastest current CPU.
I chose the fastest chip under 1k$, the i7 970, a 6-core CPU clocked at 3.2 GHz, which set me back 550$. Most reviews you will read on the net were written very shortly after it was introduced, back when it was much more expensive, and so recommend buying the 980x instead. These days however, it’s half the price of the 980x for 95% of the performance.
There is no such thing as a computer with too much RAM. Thus, I bought a 24GB Corsair RAM kit for 350$; amazingly cheap when consider that a year ago the same kits were selling for 1200$.
I’ve had good experiences with ASUS motherboards in the past, and so went with the Sabertooth X58, a cool 200$. The specs of the more expensive mobos from ASUS are very similar.
Video card, Power Supply and Case
I wanted to get in the GPU computing game, so I got a GTX 580 card, the top-of-the-line single-GPU GeForce card. The specs are pretty amazing: 512 CUDA cores and 1.5GB of RAM for 550$. It’s basically a Tesla card with half the RAM at a quarter of the price.
Two caveats about the card though. First, it’s huge. That’s why I had to buy the very large case that you see (Corsair Obsidian series). Second, it takes a lot of power. Something like 250W. Coupled with the second video card I put in there for display purposes, I went with a 850W power supply. They make modular power supplies now where the wires are removable, which is great because the interior of the computer is much cleaner and accessible now. Normally this wouldn’t be a big deal, but the case has a window on it to see inside! How awesome is that?
The case, power supply, second video card, hard drive, mounting and taxes add about 1000$, for a total of about 2500$. That’s for a computer with a top-of-the-line 6-core CPU, 24 GB of RAM, and a huge video card fit for GPU computing. All things considered, pretty cheap. I’ll post some benchmarks in a couple of days.