CUDA Quadro series comparison

I am trying to build a workstation with some GPU computing capacity. I am little lost with all the options out there. I need a little help. I will be using the computer for scientific computing. The options I have are

• 2 X NVIDIA® Quadro® M4000 de 8 Gb (Maxwell Generation 1664 x 2 cores, Compute Capability 5.2)
• 2 x NVIDIA® Quadro® K4200 de 4 Gb (Kepler Generation 1344 x 2 cores , Compute Capability 3.0)
• 1 X NVIDIA® Quadro® M5000 de 8 Gb (Maxwell Generation 2048 cores, Compute Capability 5.2 )
• 1 X NVIDIA® Quadro® K5200 de 8 Gb (Kepler Generation 2304 cores, Compute Capability 3.5)

Prices all around same. I am open to other suggestions.

Thank you

Erdem

Answer 1

[Joss Knight]

For MATLAB, unless you are doing strongly divisible operations such as Monte Carlo sampling, you are better off with one really powerful GPU than two less powerful ones since you can only use them for that kind of thing (see http://blogs.mathworks.com/loren/2013/06/24/running-monte-carlo-simulations-on-multiple-gpus/).

So you probably want the M5000 or the M5200. Choose between those based on whether you need double-precision performance or single-precision performance - the Kepler card will most likely give you far better double-precision performance (and MATLAB uses double precision by default), but if you can get by with single precision then the Maxwell architecture will give you more lastability.