Most important hardware specs?

I am buying a PC workstation to run MATLAB image processing routines on video. What PC hardware specs are the most important for maximizing the speed of my routines?

Thank you!

Answer 1

[Walter Roberson]

Up to a certain point: memory. If you do not have enough memory, then you will either fail or you will swap to disk, which is slow. However, once you have enough memory to hold your data, then additional memory only helps to the extent that your operating system uses the extra to cache files (or to run anti-virus, or to run your web browser...)

If you are reading video from files, then file access is often the limiting factor. If you are using multiple streams of live video then USB bandwidth and contention is often the limiting factor.

Once you have an image, then it depends what you are going to do with the image. There are some operations for which automatic parallelization is used, but other common operations on video turn out to be relatively serial. Using parallel computing toolbox to process multiple files automatically might or might not help, depending what is being done. Mostly using parfor on different files helps only to the extent of at most two workers per drive and sometimes only at most two workers per drive controller (beyond that you just enter into drive contention.)

For operations that are not automatically parallizable, then typically single core speed is most important than number of cores. Beyond around 6 or so cores it is not uncommon for communications overhead to start to overwhelm the benefit of adding additional cores.

If you are thinking of doing Deep Learning, then which Deep Learning can be important. Some of the Deep Learning routines benefit strongly from GPU.

[Arel Weisberg]

Thank you for taking the time to respond.

I'm not an expert at all in this area, so what you provided is very useful information.

[Walter Roberson]

A lot of the time, once you have enough memory, what you end up wanting for computations is the fastest single CPU speed, because a lot of operations end up being serial. After that then the next concern tends to be to get the fastest I/O system that you can afford -- e.g., if you are doing a lot of file access you would prefer a good quality SSD large enough to hold all the files you need for a run.

You do not need to store all of your files on SSD, just the ones you are going to work with. A moderate sized SSD with a fast hard drive would generally turn out to be more productive per dollar than a large expensive SSD. Large SSD tend to be more about convenience than about computing needs.

You would prefer at least two cores, so that the operating system can do background work and pay attention to your browser and so on while computing is going on. I have enough random things going on that I probably would not buy less than 4 cores these days.

There are programs for which having a dozen moderate speed cores is better than having four high speed cores: it depends a lot on what is being done, and on the way that the program is structured. But fewer faster cores tends to work out better in most cases involving programs that were not specially designed for multiprocessing.