Nvidia GTX 970 or 580

jump9

Reputable
Dec 8, 2014
2
0
4,510
Im building my new PC soon and have been bouncing between these two cards to use in the system. I currently have bought AMD FX 8350, Asus m5a78l-m/usb3 and 8gb crucial ballistix sports RAM and 750 watt PSU. The PC will be mostly used for rendering, 3DS Max, blender all that sort of thing, and have found that the GTX 580 is very good for these sort's of task but im not sure how the GTX 970 compares. Which one would you suggest??

My budget is under £300
 
Solution
It depends on the task. Results for AE are very different, the *only* card from the 700 series that always
beats the 580 for any test is the 780 Ti. Check reviews (I have, extensively, and I've tested a lot of cards).

For CUDA, a 980 is slightly slower than two 580s, about 10% less, and I was testing with a good 980 (1266MHz core).
Thus, a 970 will be a bit slower still.

Hence, a 970 should be faster than one 580 for CUDA, but make sure the app supports Maxwell V2 CUDA, because
not all apps do yet (AE doesn't).

Also, performance isn't purely related to absolute throughput. Indeed, one reason why the 580 is strong for CUDA
is that its high bw is shared across fewer cores, providing a lot more bw per core. Other reasons include a 2X...
I cannot imagine a single instance in which the 580 would outperform the 970..

Keep in mind the 770 is an improved 680 (and better than the 580), and the 970 is better than both of those on any day.

Theoretical performance alone it's considerably faster.

http://www.hwcompare.com/18079/geforce-gtx-580-vs-geforce-gtx-970/
 
It depends on the task. Results for AE are very different, the *only* card from the 700 series that always
beats the 580 for any test is the 780 Ti. Check reviews (I have, extensively, and I've tested a lot of cards).

For CUDA, a 980 is slightly slower than two 580s, about 10% less, and I was testing with a good 980 (1266MHz core).
Thus, a 970 will be a bit slower still.

Hence, a 970 should be faster than one 580 for CUDA, but make sure the app supports Maxwell V2 CUDA, because
not all apps do yet (AE doesn't).

Also, performance isn't purely related to absolute throughput. Indeed, one reason why the 580 is strong for CUDA
is that its high bw is shared across fewer cores, providing a lot more bw per core. Other reasons include a 2X higher
shader clock than later cards, etc.

The down side of course is power consumption, noise & heat. My 980 is very quiet, whereas my quad 3GB 580 CUDA
research box is impressively loud under load.


CraigN, don't compare or assume anything on a basis of assumptions vs. the 600 series, as the 6xx
cards were pretty terrible for CUDA. Check the AE benchmark thread by Teddy Gage, the 600s and most
700s get hammered. Of the older cards, the only one that's usefully better overall than two 580s is a
single 780 Ti (sorry Omar, you're wrong). For the newer 900 series, the problem is app support for V2 CUDA.
Check the Arion page, my quad-580 is system is no. 17, scoring 5407; by contrast, my single 980 scores 1957.
NB: system 15 on the Arion page has a typo, it should read as a triple-780. The page author hasn't fixed it yet.

Btw, Teddy's very first post on the AE thread shows his 580 beating his 680.

Note that for gaming of course, a 970 will leave a 580 in the dust, and indeed in many cases will likely be
faster than two 580s. I've done a fair few tests, comparing a 980 to one, two and three 580s, results on my site.
Check the Unigine Heaven/Valley tests, CoJ and Stalker/FC2 at high-res.

Ian.

PS. Don't take my word on all this, check toms' own review:

http://www.tomshardware.com/reviews/geforce-gtx-760-review-gk104,3542-19.html

PPS. To the original poster, a 970 will be faster than one 580, but will definitely not be
faster than two 580s. Thus, it depends how much you care about noise, power, heat, etc.,
as the 970 will certainly be better for environmental factors. Actually, 300 UKP should be
enough for three 1.5GB 580s, or at least two 3GB 580s these days.


 
Solution


mapesdhs,

Thank you for your polite reply and also for thoroughly explaining why. I appreciate that. I will add that thread to my favorites so I can better advise in the future. I appreciate all the information your provided! :)

To clarify, I did more research after and did explicitly say I believed the 970 would beat one 580 though obviously not by a large margin, which seems to agree with you, but it definitely looks like multiple 580s does extremely well very quickly by comparison to other paired setups. This is good to know :)
 
Yep, you're right, though after doing so much testing (I've meddled with about 15 different 580s now), I would
definitely warn anyone considering using 580s that they can be pretty noisy under a full render load, especially
any reference 580 with a standard cooler. Also, good air flow is essential, doubly so with a mbd that does
not have tri-slot spacing. Then there's power consumption and the effects of heat on other parts of the system
(most 580s dump some or most hot air inside the case), which can have some weird effects if one is trying to
use a water cooler, as I was with my HAF 932 case that has a 3930K + H110. The biggest advantages
of 580 cards is their lower upfront cost; for some, this may be more important (certainly was for my gaming PC,
which had a 580 SLI config for about a year). My gaming PC had two reference 3GB 580s (Palit @ 783MHz),
whereas the AE box has four good 580s, all MSI 3GB Ligtning Xtremes (832MHz, but they oc to 950 or 1GHz).

Btw, Arion as a render engine scales better than most CUDA-based systems, but sadly it's not typical
of many apps that use CUDA, so its a bad idea to use Arion results as a means of judging how a GPU
config will behave with some other app, because Arion is at the top of the pile of performance exploitation.
I had a chat with one of the Arion staff about this, he said they put a lot of work into writing code that
can make good use of different models of GPU being used in the same system, good parallelism, etc.

This is why, for AE, a 580 easily beats a 680, but for Arion it's the other way round. Different code/app.

Ian.

PS. If it's of any help, I found a stable 900MHz oc for a Palit 3GB 580 was obtained with 1075mV, using
1800MHz shader and 2150MHz RAM.