Balence between CUDA Cores and Clock Speed

Dethan1999

Reputable
Oct 11, 2015
198
0
4,760
Hey everyone. I've always been curious as to what is more important to a GPU: CUDA cores or Speed.

I noticed earlier that the GTX 660 has 960 CUDA cores, which is far more than my GTX 750 TI. That itself confused me because I thought lower series has much lower CUDA cores and speeds. Anyways, even though the GTX 660 has more CUDA cores than the GTX 750 TI, it has a lower clock speed than the 750 TI. That got me wondering, what is more important? And what has more of an impact on overall performance? I also noticed that one of the best GPU's out there happens to have a base speed of 705 mhz (Titan Z). But with a somewhat lower clock speed, it comes with a whopping 5760 CUDA cores! That leads me to think that CUDA cores have more of an impact, but what are some flaws of having more CUDA cores with a lower speed? Do GPU's with more cores and lower speed get work done more efficiently but slowly? How does that all play out? And is there a balance between the two specs that create the "ideal" GPU?
 
Solution
TL;DR:
I use this as a general guideline for how one card compares to another:
http://international.download.nvidia.com/webassets/en_US/shared/images/products/shared/lineup-full.png


To answer your question, generally parallel processing will trump speed in most cases, so more CUDA cores would be the advantage. There are other things to consider however, memory speed, bandwidth, core speeds, etc..., they all have to work together to get the performance you need.

However since card manufacturers try very hard to make all these things 'even' so you don't have any significant bottlenecks it mainly boils down to over # of cores.

I just use the picture, find a card I like and then read up on some reviews to make sure they hype is...

Kurdain1

Distinguished
Nov 30, 2007
154
2
18,715
TL;DR:
I use this as a general guideline for how one card compares to another:
http://international.download.nvidia.com/webassets/en_US/shared/images/products/shared/lineup-full.png


To answer your question, generally parallel processing will trump speed in most cases, so more CUDA cores would be the advantage. There are other things to consider however, memory speed, bandwidth, core speeds, etc..., they all have to work together to get the performance you need.

However since card manufacturers try very hard to make all these things 'even' so you don't have any significant bottlenecks it mainly boils down to over # of cores.

I just use the picture, find a card I like and then read up on some reviews to make sure they hype is accurate.

Hope this helps.
 
Solution

Dethan1999

Reputable
Oct 11, 2015
198
0
4,760


Good to know! That answers a question I've had for quite a while. I appreciate the help!