I am looking for a high-performance computer to run scientific simulations on. I will be taking this computer with me on business trips. It does not need to be a laptop, but it needs to fit on an airplane and in the back of a rental car without too much trouble. It can be checked baggage on an airline. It does not need a screen, and I am not interested in any specifications other than total CPU performance. The application I’m using can be considered to scale linearly with the number of cores.
Something with 12 Xeon-class x86 cores is what I’m looking for. Conventional server options from Dell and the like are big, heavy things designed to hold many hard drives and remain immobile in a rack for years. “Blade” servers are compact and dense, but require a huge, heavy backplane (i.e. they are not that dense if you only need a few). I cannot use GPUs. Big laptops max out at 4 cores.
Any suggestions? The computer has to be physically with me, it cannot be hosted somewhere and accessed remotely.
Your requirements are weird. You could save a ton of money by getting rid of your “computer can’t be accessed remotely” requirement-- just rent the CPU power from Amazon for the few minutes you need it, save tons of money.
To actually address the question, you can cram a lot of CPU/GPU power into a full tower case, and they’re still relatively common and easy-to-find despite mid-towers taking over the world. Check NewEgg. Server cases are usually designed to rack-mount and as a result are heavy and awkward to transport.
If you’re not experienced with building computers, you should find a local computer shop or expert to advise you.
And in the mean-time, get whoever wrote your app to start porting it to use CUDA or another GPU library, because you could get a crapload more computer power that way.
I took a look at Dell’s site and you can build a PowerEdge T620 with dual Xeon E5-2650, giving a total of 12 cores at 2.0 ghz for about $4000. It’s a “server tower” case, which is big and heavy but a lot easier to move around than a rackmount case. Dell lets you go as high as a E5-2690 2.9 ghz in that case, but of course the cost skyrockets.
If you need more than that, you’re going to have to have someone build you something custom, I believe.
We have several $5000 T620’s already. They are incredibly heavy and difficult to transport. A full-tower case is just as bad, and even a mid-tower would have a huge amount of wasted space. I don’t need a big graphics card, nor an optical drive, nor 5 3.5" drive bays, etc. I need a bunch of CPU, an ethernet jack, a VGA port, and some kind of keyboard. That’s it.
GPU are only effective for certain kinds of computation. This is not one of them.
… so why not go to the source of the issue and solve the problem of why the hardware needs to travel in the first place? Is the implied question there I guess. Or rewrite the software to use supplement the CPU with GPU power, and you might be able to get by with one of those laptops you’ve looked into. EDIT: oops, just noticed you said GPUs won’t help much.
I doubt you’ll find a better solution than the Dell server boxes you’re already buying. You may consider hiring someone to see if they could pack those same components into a smaller case, but I doubt they’d be able to do better than Dell… power and cooling requirements don’t give you much flexibility, the PSU required alone is pretty hefty. The “empty space” of the case is required to ensure proper airflow, Dell’s not making the case that size just to piss you off.
I agree with the others. Get rid of the portability requirement. Why bother with something that will need to be checked baggage? There’s a good chance your machine won’t be available when you need it that way, either in the wrong location or not working. You’d also be dependent on the electrical service available on the road, so unless you can tolerate sudden power-downs then you’ll need battery backup on top of that. Even if remote operation is impractical you can have a machine in a fixed location with all the CPU you need, and on the road use a more convenient high end laptop that just runs slower while you are away.
The problem is insoluble. The servers cannot be connected to the internet for security and legal reasons.
Well, looking at the T620 case, the dimensions are not set by the size of the heatsinks and fans, but by the large storage backplane that supports something like 12 hard drives, as well as various other types of expansion slot that are currently empty. The power supply is very compact, and the case has an empty slot for a redundant one. The airflow is all forced air with a shroud over the heatsinks, there is no free circulation at all and all the empty space is physically separate from the cooling airflow. The motherboard supports triple-redundant RAM or something, so there’s a huge number of unused RAM slots. There are just tons and tons of ways you could shrink the case.
Well there you go, you’ve answered your own question. Hire an engineer to create a reduced case for you. It’ll be expensive of course, but your company’s already paying 10 times what it should be paying due to “security and legal reasons”. (Which is ridiculous BTW; AWS doesn’t prevent you from using strong encryption and your legal department should be clued-in to 21st century technology. You’re much more likely to get a physical server stolen from airport checked baggage than you are to get data stolen from a cloud computing provider.)
Sorry I’m not trying to be snarky. The problem here is that you’re looking for a product that has virtually zero demand. If you’re wondering why nobody builds it, it’s simply because nobody needs it. So you either make-do with what you can find, or build it yourself.
You can configure a six core i7 from Shuttle. That has to be one of the smaller packages you are going to get. Scroll down to the bottom to see the dimensions.
Actually, this is a response to 21st century technology. Most companies just follow industry best practices and if they have a security breach, they say “Oh well, we did the best we could,” and have their asses sufficiently covered. But if you genuinely cannot leak sensitive data, modern computer security is a joke, as evidenced by the numerous security breaches at ostensibly competent companies every week. The only foolproof security solution is to simply disconnect the computer from the internet (and not connect any strange USB sticks). Then, if someone wants your data, they actually have to send a real person after it, and that person can be caught, tracked, etc.
Encryption is useful for securing data being stored or in transit, but is totally useless at securing data that is actively being processed. If the CPU can access the data to process it, then so can anyone who compromises the machine. Virtual machine systems have an additional set of security issues to worry about.
Your last statement is completely and laughably false. Nevertheless, we keep the hard drives in our carry-ons, and they do use whole-disk encryption, so even if one were stolen in transit it wouldn’t be a huge deal.
And these servers get pretty regular, heavy use, certainly not just a few minutes per day. AWS is actually not that cheap for sustained use.
Hah, I had not even considered the Mini. Dealing with three computers instead of one would be a little cumbersome but it is undeniably compact and portable. Unfortunately, I don’t think it uses server-class processors. But that may well be the answer.
You said it scales linearly with cores which means interconnect is not a bottleneck, so you could get a few of the most powerful laptops on the market and network them.
Curiously this is a problem that I face from time to time, usually in the abstract, but I like to keep abreast of the issue.
You really need to get some much more specific answers about the nature of your problem. OK, it is embarrassingly parallel, but not pure numeric to the point a GPU or DSP solution will work. The critical question will be cache sensitivity, and then memory latency sensitivity - which are of course interrelated. I/O needs cannot be ignored, but it sounds as if you know that they are not a problem. This gives you the sweet point where you balance core architectures - fewer, faster, more cache, versus more, slower, less cache. The difference this makes for an individual application can be significant, and may make or break one solution over another.
At the lots and slower end of the scale, I would seriously want to know if a Xeon Phi card is capable of handling your problem. This packs an astounding amount of x86 capability in a very small (if rather power hungry) PCI card. Not cheap, but very very interesting. Intel PDF. Tom’s Hardware article.