Tesla K20 Capabalities

Hello,
I am running matlab simulation that requires matrix inversion of 14400 on 14400.
I would like to know the following please:

  1. Will Tesla K20 Memory (5GB) will be enough to store this matrix ?
  2. What is the expected acceleration using Tesla Vs i7-3280, CPU @ 3.60GHz?
  3. Do you think that Quadro K6000 that have 12GB will be more suitable due to this memory size in this case?

I would really appreciate any information,
Best Regards,
Tamir

A matrix of 14400 by 14400 will fit in the 5 GB. If you have double multiply by 8 bytes for each point you get 1.6 GB. What are you using it for? If you use cufft you are at the limit, because the cufft library allocates a part of ram for intermediate operations. It might be enough or not depending on what are the other things needed in addition to the main matrix.

Pasoleatis, thanks for you answer.
I am using this inversation to solve optimization problem.
I am not using FFT in the code. according to what I understand if I clean the memory before this inversion command so there is no concern of Out Of Memory.

Can you give me information regarding my other questions ?
And also, can I install this card on regular desktop with PCI-e ?

Thanks a lot!
Tamir

  1. Yes. But take also into account the additional dependencies which depend on the problem. Multiply the total numebr of points with 8 bytes (4 for float or integer variables) and then divide by 1024^2 you get 1.6 GB.

  2. No idea.

  3. I do not think it is needed 12 GB. Depends on 1).

You need an actively-cooled K20, known as K20c, for that.