6
$\begingroup$

I am looking at the Nvidia GT-860M which comes in both the old Kepler and new Maxwell architecture.

The old one (1152) seems to have almost twice the cores as the new one (640). The new one has a compute capability score of 5.0 compared to the old one of 3.0. Yet, the new cores are more powerful, according to:

https://devblogs.nvidia.com/parallelforall/5-things-you-should-know-about-new-maxwell-gpu-architecture/

Which is better for computation? E.g. fitting random forest models?

$\endgroup$

2 Answers 2

4
$\begingroup$

While the Kepler card has more cores, the Maxwell ones run at a higher clock speed. This review actually compares the two, although mainly for gaming purposes, and the Maxwell has the Kepler beat in almost all cases. The margins are relatively small however. Looking at the numbers, I would hazard the guess that the Maxwell card will perform better for your purposes, particularly if you make sure to compile your GPU code for that compute model (5.0).

Keep in mind that desktop GPUs are always quite a bit faster than their laptop equivalents (i.e. 780M is a lot slower than the 780), so if a desktop computer works for you that will give you better performance.

$\endgroup$
2
  • $\begingroup$ Thanks for your answer. However, if the routine is highly parallelisable,isn't it better to have more cores than less? $\endgroup$
    – Alex
    Jun 24, 2014 at 0:30
  • 2
    $\begingroup$ Not as such, no. Simply put, all you want is for the combined power of the cores to be as high as possible. There are a lot of other factors as well however. Many (most?) real world GPU programs are limited by memory bandwith, so anything that hides memory loads will help. This includes the amount of L1 cache (of which the Maxwell has more), the number of registers, etc. $\endgroup$
    – LKlevin
    Jun 24, 2014 at 8:42
0
$\begingroup$

As an important factor for scientific computing which has not been mentioned in @LKevin's answer, let me point out the difference in double-precision performance between the Kepler and Maxwell architectures. As Wikipedia says:

The theoretical double-precision processing power of a Maxwell GPU is 1/32 of the single precision performance (which has been noted as being very low compared to the previous generation Kepler).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.