Update: Added more Tests and Results - Hitman: Absolution Benchmark, Unigine: Valley Benchmark, Left 4 Dead 2 Unscripted Gameplay Test, Tribes: Ascend Gameplay Test and Rollercoaster Tycoon 3: Platinum Gameplay Test
Recently becoming the proud owner of another NVIDIA card, I swapped out my trusty AMD/ATi Radeon HD 6870 from ASUS that has served me so well and lovingly joined together two GTX 560 Ti videocards in electric matrimony to try out some SLI. I share now my experiences, the results of many tests and answers to questions I have seen in many places online, including one I had myself.
This is also a 'Quick' Edition, a version of the testing and results that are text-only for now, where I may do a full version as time permits in the future, with screen recordings, graphs and other materials that I would like to share and would help people out.
The Experiment
For this TestRun, with two NVIDIA videocards, I try answer the question: "Does SLI really make a difference?" and also "Where to put PhysX?". Sure, the answer of 'if it makes a difference' is answered other places online, with pretty graphs and numbers; but I am one of those people that, although accepting, don't 'truly believe' things unless I can test it out myself. This is a human stubbornness I know, but it also helps me explore many things I would not have otherwise experienced and allows me to share those results and experiences with others. "Does it really double performance?" is probably a more specific question of what I'm feeling, as of course it should make a difference, after all I am putting another 384 processing cores towards throwing around colorful shapes on my screen. I want to know if it even 'doubles' the awesomeness of games that I play. There are two cards in there now, so what's the difference in performance?
Another aspect I plan to explore is regarding PhysX, NVIDIA's lovechild of bouncing balls and splattering sparks. Does it affect game performance? Which card should run it if I have it on? Is it better to run it on a dedicated card [that only uses it for that] or off of the CPU? These are my own questions as well since finding a second NVIDIA gpu on sale... PhysX is only in a hamfisted-handful of games - and is only going to be found on NVIDIA GPUs, but it is supposed to improve visual/perceived quality in the games that do use it, due to the fancy waving flags, shattering glass and splattering blood [and the like]. I am going to test the options of having it on the first or second SLI card and how PHysX runs off of the CPU, too.
The Test
For this TestRun, I used utilities such as FurMark and FluidMark, Unigine's Heaven Benchmark and a handful of games to benchmark and test with, such as Just Cause 2, Alien Vs. Predator, Lost Planet 2, Battlefield 3, Unreal Tournament 3 and Batman: Arkham City (all but the first two utilize PhysX). I am currently running an AMD 6-Core CPU and two GTX560Ti's in SLI mode on a GIGABYTE 990FXA chipset mainboard.
I assume this is still an average-to-above-average system at the time of this writing (perhaps upper midrange but definitely not top-of-the-line) and it would help out anyone with a similar rig in the future as well with examples of some of the performance difference they will see [if they add a second videocard].
The Data
Here are my benchmarking test results and average framerates, comparing SLI mode with a single GPU:
FurMark: Score (2127), Average Framerate (40fps)
FurMark: Score (3959), Average Framerate (65fps) SLI
Lost Planet 2 - Benchmark B: Average Framerate (44fps)
Lost Planet 2 - Benchmark B: Average Framerate (103fps) SLI
Just Cause 2: Concrete Jungle Benchmark, High Settings, Average Framerate (28fps)
Just Cause 2: Concrete Jungle Benchmark, High Settings, Average Framerate (51fps) SLI
Alien Vs. Predator: Benchmark Run, 1920x1200, x16 Filtering, Average Framerate (44fps)
Alien Vs. Predator: Benchmark Run, 1920x1200, x16 Filtering, Average Framerate (90fps) SLI
Unigine Heaven Benchmark: 1080p, x4x4 Filtering, Score (960), Average Framerate (38fps)
Unigine Heaven Benchmark: 1080p, x4x4 Filtering, Score (1717), Average Framerate (68fps) SLI
| Unigine
Valley Benchmark: 1080p, Ultra Settings, 8xMSAA - Score (1077), Average
Framerate (26fps) |
Unigine
Valley Benchmark: 1080p, Ultra Settings, 8xMSAA - Score (1741), Average
Framerate (42fps) SLI
Hitman: Absolution Benchmark: 1080p, Ultra Settings, Average Framerate (32fps)
Hitman: Absolution Benchmark: 1080p, Ultra Settings, Average Framerate (33fps) SLI
(Hitman: Absolution does not appear to take advantage of SLI and/or is not Optimized for it, despite double-checking that SLI mode is enabled and trying to utilize NVIDIA's 'Custom Game Profile' for it, as well - the following two games also do not appear to take advantage of SLI mode)
| Tribes: Ascend - 1920x1080, Very High Settings, Average Framerate (51fps) |
| Tribes: Ascend - 1920x1080, Very High Settings, Average Framerate (53fps) SLI |
| Rollercoaster Tycoon 3: Platinum - 1600x900, 8xAA, Reflections and other (17fps) |
Rollercoaster Tycoon 3: Platinum - 1600x900, 8xAA, Reflections and other (19fps) SLI
|
Here are some unscripted, raw gameplay framerates, averaged:
(multiple samples taken of each with FRAPS' built-in benchmarking utility)
Alien Vs. Predator: Marine Storyline, Average Framerate (47fps)
Alien Vs. Predator: Marine Storyline, Average Framerate (89fps) SLI
Battlefield 3 : Multiplayer with 64 Players, Average Framerate (33fps)
Battlefield 3 : Multiplayer with 64 Players, Average Framerate (61fps) SLI
| Left 4
Dead 2 - 1920x1080, Very High Settings, 4xMSAA, 4xAF, Extra Film Grain
(128fps) |
| Left 4
Dead 2 - 1920x1080, Very High Settings, 4xMSAA, 4xAF, Extra Film Grain
(185fps) SLI |
*Planetside 2 : Multiplayer with many players, BioLab, Average Framerate (36fps)
*Planetside 2 : Multiplayer with many players, BioLab, Average Framerate (37fps) SLI
[*Game does not appear to be fully optimized currently, article with more detail here:
http://gametipsandmore.blogspot.ca/2012/11/planetside-2-and-ultra-settings-quality.html)
Here are the PhysX tests, done with FluidMark:
FluidMark: Score (5338) single-card mode, PhysX on the card, Average Framerate (87fps)
FluidMark: Score (5449) SLI, PhysX set to (GPU1), Average Framerate (89fps)
FluidMark: Score (7512) SLI, PhysX auto-selected (GPU2), Average Framerate (123fps)
FluidMark: Score (7379) SLI, PhysX set on (GPU2) Dedicated, Average Framerate (121fps)
FluidMark: Score (1291) SLI, PhysX set on (CPU), Average Framerate (21fps)
The Conclusion
So, does putting in a second videocard practically 'double' your gaming performance? Dang right it does. Now, I have read that it 'should' in many other articles online, but again, I had to test it out for myself. Almost across the board, in every game I threw at it, the performance was almost doubled. Wonderful stuff.
For those about to spend money on a second GPU for SLI because now they know it 'will definitely make a difference', I should mention as a Tip here that the more cards you add in (for 3-way or 4x SLI or Crossfire, for instance) the lower the rate-of-return you actually get on the investment.
In other words, adding a third card won't 'double the performance again' and adding a fourth card won't 'double that performance yet again' - it won't even give you 300% and 400% performance [for the third and fourth videocards added]. You will actually see diminishing returns as you add more cards.
Still, if you have enough money for only a second GPU that matches your 'old' one, but not enough money to splurge on a new, faster one, clearly SLI/Crossfire is the way to go, it really does make a huge difference.
The PhysX tests are interesting, as they answer a question that I myself was wondering: does it matter whether you set PhysX on GPU1, GPU2 or the CPU, and should a GPU be dedicated to it?
It seems to make sense, that although a powerful, complex 6-core processor, my CPU is nowhere near the raw number-crunching [if 'simpler'] streamlined data-geyser that the 384-core video card is. The difference can be clearly seen in the results. On the CPU, PhysX calculations [alone/isolated] gave a 23fps output of gushing blood on the screen, while even just one videocard running PhysX by itself put out 87fps of crimson data. Adding another card and telling that GPU instead [of utilizing the first card] to 'calculate the physics if you please' took the framerate up to over 120fps. Clearly, NVIDIA is showing that the videocards were made to handle the floating leaves, sparks and bouncing balls of PhysX ...and ball handle they did.
What is interesting is the difference that was made, depending on which card you chose to handle PhysX. When the PhysX was set to GPU2 (the card not plugged directly into the monitor in my case), the highest score was achieved. When the PhysX was set to GPU1 (the card plugged directly into the monitor), it scored almost exactly the same as if there was no second card in the system at all. Now, the utility (FluidMark) does test 'only' PhysX and there isn't much to see on the screen other than some bouncing baubles, but it seems to show that if you put PhysX on the same card that you are using as your display card (that your monitor plugs into), then it might in fact affect performance of that card to display the pretty images on your screen slightly - or at least, to 'deal with PhysX and show pretty colors on your screen' at the same time.
Either way, they are both better than utilizing the CPU, as the pitiful performance when on the CPU (at least for 'mainly PhysX only' calculations) is over 500% slower than the performance seen when PhysX calculations are put on the second video card, and still calculated at over 400% slower compared to when PhysX is put on the first video card in an SLI setup. Clearly it is better to have PhysX calculations done on a GPU - and perhaps not put it on the main display card if you can.
Testing the ramification of this in actual gameplay, I played some Unreal Tournament 3 and Batman: Arkham City, while juggling the PhysX of both games to be calculated on the CPU, GPU2 and GPU1 (the main display card where the monitor is plugged in). Using FRAPS for collecting benchmark data in Average Frames Per Second, I saw no more than a few frames difference between all modes.
That is, playing with PhysX set to GPU2, PhysX set on GPU1 or putting the PhysX onto the CPU, I saw only a few frames of difference in actual gameplay performance. Mind you, it can be argued that I didn't see as many sparks, blood, snow or leaves float across my screen at times; but overall, if you are concerned with only "the game performance ramifications" of where to put PhysX, The Answer Is: playing a game, it doesn't really matter (but the CPU calculates PhysX slower than putting it on any card, as seen in the data).
Overall, if you can't afford that fast, newly-released videocard you want, but you can afford to get another GPU of the same type as the one you have now - and your motherboard and powersupply can support it - it's totally worth it to get a second videocard for SLI/Crossfire performance (and if you're going the NVIDIA route, it doesn't really matter where you put PhysX).
Have fun upgrading and See You In The Games If The World Doesn't End Soon!
|
+++-GTAM.jpg)
+++-GTAM.jpg)