The comparison no one asked for but here it is. I’ve always been fascinated by nVidia’s SLi technology, shotgunning multiple videocards to multiply graphics performance. Of course, it’s not a perfect technology. You’ll see the same or better performance investing the same money into a single better graphics card, performance doesn’t scale 1:1, and the difference between minimum and average frame rate typically grows as you add cards (citation). So why do it? For me, I was adding a second card after I had owned the first one for a while. Prices dropped and and it became reasonable to invest a few extra dollars to increase my system’s performance. Later I was simply given two identical cards that were a whole generation newer, so I had no real reason to run a single card. Now that I’ve been gifted a GTX 970 SC it’s time to compare multiple older cards to a single current one.
The system the cards are being tested in consists of a water-cooled AMD FX-8350 overclocked to 4.415 GHz, 16 GB of DDR3 at 2427.9 MHz, and an ASUS M5A99X-EVO motherboard packed inside an NXZT H440 case with stock cooling. The dual GTX 560 Ti cards are from Palit (I hadn’t heard of them either) while the GTX 970 SC is from EVGA.
First up is the most standard test, 3D Mark Firestrike. Thanks to consistent settings across all systems this is the most reliable test for comparing performance, though it’s arguably not “real-world” performance. The highest score I was able to manage utilizing the dual GTX 560 Ti configuration was 5332. With the single GTX 970 SC my score jumped 62.9% to 8685. While this is a pretty huge jump I wasn’t completely satisfied since my score was still slightly below the “Oculus Rift spec” of 9271. To combat this I moved the card up to the top PCI-e slot, figuring the lower slots might be limited to x8 speed. Moving the card up brought my Firestrike score up to 8854. This could be because of the added bandwidth afforded by the x16 slot or it could be the result of, well, any number of things. I didn’t move the card down to see if that brought the scores back down, but it was at least an interesting change.
Next was the 3D Mark Sky Diver test. I scored 16,082 with the dual 560s and 24,612 with the single 970; a gain of 53%. It makes sense that the jump here wouldn’t be as big as we saw with Firestrike since it’s using older rendering technology and throwing memory and DirectX 12 support at it doesn’t really help here.
The most impressive by far test was Tomb Raider. Using the default settings on “ultra”, a resolution of 1680×1050 @ 120 Hz and V-sync off I got a minimum frame rate of 90, maximum of 140, and an average of 116.4 frames per second. With the new setup those figures increased to 130.1, 198, and 165.7 frames per second for a gain of 44.6%, 41.4%, and 42.4% respectively. That in itself isn’t impressive but when I changed the settings to “ultimate” (identical to “ultra” but enables TressFX for rendering hair as strands rather than a single mass) the frame rates fell through the floor on the old setup. Minimum, maximum, and average frame rates were 3.6, 35, and 13. The minimum frame rate was under four frames per second and the average a meager 13. With the new system we saw a minimum of 82, maximum of 132, and an average of 107.9 frames per second. That equates to a gain of 2,177.8%, 277.1%, and 730%.
One benchmark came back with kind of the opposite result of Tomb Raider. Cinebench’s OpenGL test came back with 89.57 points with the 560s and 82.32 points with the 970, a gain of only 3.1%. I also ran the CPU test, just for giggles, and that resulted in a gain of 4.7%. Initially I thought that maybe the test only ran on the CPU using software but the reporting in Cinebench also tells you what GPU is being used so I’ll have to investigate.
Going back to trying to match or surpass the Oculus Rift spec score, when I ran Firestrike Extreme and Firestrike Ultra (neither of which could run on the 560s due to only having 1 GB of memory) I somehow got better results. In Extreme I beat the Oculus Rift score 5,136 to 4,926, and in Ultra I beat it 2,746 to 2,596. Other users are supposedly getting over 10,000 points in the standard Firestrike test using the same CPU and similar GPUs so I think there’s still some work to do.
Another observation, outside of performance, is noise. My 560s and my 460s before that both started to sound like aircraft taking off when they started to sweat, even after cleaning and reapplying thermal compound. I’ve supposedly thrown some pretty demanding processes at this new 970 and it never seems to break a sweat. In fact I can never tell if the fans are even spinning (which they don’t at idle). After running a benchmark the temperature drops from an already-cool low-70s to low 50s within seconds. I’m honestly baffled at how this is even possible but I’m not complaining; I’m just not sure if the 970 is even being taxed at this point.
Conclusion: Shockingly, a single modern, high-end video card is better than two old ones taped together.