It could, but it doesn’t have to. The key is buying as much machine as you need, and no more.
The entry-level cost for some of the commercial turn-key clusters today is in the $10,000 range. This will get you 8-12 cores, the interconnect, disks, and operating system you need to have a small but functioning cluster. Prices go up from there; a 64-core Intel Xeon system today with a reasonable amount of memory and disk will cost you in the $50,000 range. But remember that prices change all the time, and usually go down, so you’ll want to check with a least a few vendors to make sure you are getting a good deal.
But how much do I need?
This is a key question, and one you’ve got to have the answer to before you buy to prevent buying more than you need or getting locked into a system that is too small to solve the real problem you are interested in.
The best way to determine how large your high performance computer should be (how many processors, what kind of network, how much memory, and so on; see this article for a description of the elements of a cluster) is to run your application on hardware similar to what you are considering and see if it meets your needs. Many vendors have already done this for a pretty big set of applications and problems, and if you check around you may find that someone already has a case study that they can share with you to help inform your decision (you should also check the HPC section of their website; some vendors post quite a lot of information online that can help). If they don’t already have exactly what you need, many of them maintain clusters on which customers can run their own benchmarks and “try before they buy.”
You can also search the internet for people solving similar problems to yours to find out what hardware they chose, and how well it worked for them. A good source of this kind of information is the case studies published in insideHPC’s HPC and your Business portal where you found this article.
One thing that you should avoid doing is buying on the datasheet alone. Every hardware manufacturer publishes specs on their solution that tell you what, in theory, hardware is capable of doing. These numbers are always best case, and the best case almost never happens in the real world. If you can’t run your own application on the specific problem you are interested in solving, or find someone who has already run it, then look for published benchmark data on other applications and problems like yours. Never make your decision on the theoretical flops (floating point operations per second) alone.