Review Index:
Feedback

NVIDIA GTX 980 3-Way and 4-Way SLI Performance

Author: Ryan Shrout
Manufacturer: NVIDIA

Power Consumption, Sound and Closing Thoughts

Okay, we have looked at performance from a bunch of different angles, now let's compare power and sound levels of 4-Way SLI.

View Full Size

Our power testing was taken from Crysis 3; we obviously had to pick a game that had good scaling up through the 4-Way SLI results in order to make sure each GPU was being utilized as much as possible. First, the most interesting part of this for me is that even with a set of FOUR GeForce GTX 980 cards running simultaneously, this system only draw 689 watts at the wall! The idea that you could likely run a 4-Way GTX 980 system on an 850 watt power supply is mind blowing considering we needed 1200-1500 watt units for quad-GPU setups no that long ago. If you look back at my dual Radeon R9 295X2 Quad CrossFire story you'll see that system used 1261 watts at the wall - 83% more power than these four Maxwell GPUs at work.

View Full Size

Sound results are interesting as well, and could be confusing without some explanation. First, notice that the 2-Way SLI sound levels are only slightly louder than a single GPU, yet the 3-Way and 4-Way sound levels are MUCH higher. How can that be? Well, in our 2-Way configuration, the GTX 980 cards have a two-slot space of empty area between them, allowing both card to draw in cooler air much more easily. When you stack three or four cards right next to each other, the fans (on all but the bottom card) need to work a lot hard to bring in enough air to keep the GPUs at their expected temperatures. That means the fans much spin faster, generating more noise.

Remember that the GTX 980 reference cards had a partially removable back plate to make more room for air flow? As it turns out, the noise levels actually WENT UP in my testing when we took those off? I can't say for sure why but it appears that with those panels on, the air flow was "smoother" into the fan, generating less turbulence, and thus, noise.

View Full Size

Here is a comparison of fan speed and GPU temperature measured with both the back panel plate on (closed) and removed (open).

View Full Size

Fan speeds and temperatures were measured on the 2nd graphics card, sandwiched between the 1st and 3rd cards, the card that has the most issue remaining cool and calm under gaming pressure. You can see in the graph above that with the back plate open (removed) the fan is spinning around 400 RPM slower than with the back plate closed off.

View Full Size

In terms of real world temperature differences between an open and closed GTX 980 back plate, it is measurable. With the back plate on, even with the fan spinning 400 RPM faster (as seen in the graphic above) the GPU remains 3-4C hotter than with the back plate portion removed.

Looks like the designers were indeed onto something when they built in those removable back plates!

Closing Thoughts

Multi-GPU gaming has always, and will likely always, present more problems to gamers than any single GPU solution will. With a single, fast GPU you will nearly always get 100% of your paid-for performance while going down the road of SLI or CrossFire always means there will be some sacrifice of individual GPU performance for the scalability you get. Our testing today with a set of four GeForce GTX 980 4GB reference cards proves that this is still indeed the case, despite claims from all parties involved that it would be better with this generation. Maybe it still can be, but drivers need to be optimized clearly.

View Full Size

Remember those graphs I showed you on the first page of this story? Now we have added to the "optimal" line the games we tested here today at the 4K resolution. The closer each line is to the blue line, the better the scaling of SLI in that particular title. For example, both Sniper Elite 3 and Crysis 3 do very well at 2-Way, 3-Way and 4-Way SLI. However, a game like Metro: Last Light or even Bioshock Infinite shows very little scaling beyond 2-card configurations.

Nearly all of the games we tested were able to show relevant and useful scaling with two GPUs and do so without a major increase in frame time variance that would lead to a stuttery, messy gaming experience. Nothing's perfect, and games like Bioshock Infinite and Skyrim don't give you the extra performance that you might expect out of the box. But more demanding games like Crysis 3, Metro: Last Light and even Battlefield 4 proved that dual GTX 980s can work very well and provided added performance for high resolution displays or gamers that want to hit peak frame rates.

View Full Size

NVIDIA's fancy SLI bridges only support 3-Way SLI

But when we stretched to 3-Way and 4-Way SLI, only Crysis 3 and Sniper Elite 3 really provided any indication that using more than the two graphics cards was a worthwhile venture. Those two games showed expected scaling rates of 90%, 39% and 29% as you go up the multi-GPU ladder. That means for a second investment of $550 (the cost of another GTX 980) you will get 90% added performance. Another $550 investment nets you another 39%, while a fourth payment of $550 only gets you 29% more. Even when 4-Way SLI works well, the value proposition is tough to argue for.

But I guess that was never the intent. 4-Way SLI users are a rare group and I would argue that they pile on the GPUs more for show and the occasional kick ass gaming performance rather than for constant, improved gameplay. At least I hope that's why they do it, because otherwise they are kidding themselves.

View Full Size

I can't deny that seeing a stack of four GTX 980 cards in a system makes me anything but nervously excited, I just don't think I would recommend actually shell out the money for a configuration of their own.


October 10, 2014 | 02:28 PM - Posted by AJ (not verified)

It would be awesome to see this run again with water blocks and see what the difference might be.

October 10, 2014 | 05:18 PM - Posted by Anonymous (not verified)

Overclocking the gpus is not going to change the scaling factor.

October 10, 2014 | 02:39 PM - Posted by Anonymous (not verified)

Maybe you didn't get full utilization out of 4-way due to Maxwell asking for more power, there for limiting itself.

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-980-970-maxwell,3...

If each card is capable of asking for 290watts then a 1200watt PSU will not be enough.

http://media.bestofmicro.com/J/E/454298/gallery/00-Power-Consumption-1-M...

Should have tested with a 1500watt+ PSU see if that shows any performance increase.

October 10, 2014 | 03:11 PM - Posted by Ryan Shrout

I had a power supply runnig at 1200 watts, so I am sure this isn't a problem of power delivery.

But good link!

October 10, 2014 | 04:47 PM - Posted by Anonymous (not verified)

92% efficiency on a Corsair AX1200i is 1104watts. If the GTX 980 are asking for 290watts x 4 a total of 1160watts. That leaves 40watts for the rest of the system.

I don't think 40watts can be enough for the rest of the system when the i7 3690X idles at 99watts and during load uses 216watts ? That puts the system 1380watts just by CPU+GPUx4 alone well over 1200watts.

October 10, 2014 | 05:11 PM - Posted by Anonymous (not verified)

Power supplies do not work that way. They're rated for DC output wattage.

October 11, 2014 | 01:45 AM - Posted by SeaJay (not verified)

Firstly, 1200W is what the PSU supplies to the system. Efficiency doesn't affect the output wattage, it affects what the PSU draws from the wall.

eg. if it's supplying 700W, the AX1200i has an efficiency of ~92% so it draws 700/0.92 = 760W from the wall.

eg2. if it's supplying 1200W, the AX1200i has an efficiency of ~88% so it draws 1200/0.88 = 1360W from the wall.

Secondly, the 980 don't run at a continuous 290W draw, it has occasional current spikes. If the average system draw is less than 700W, a 1200W power supply should have been perfectly fine to cover that sort of thing.

October 11, 2014 | 08:19 PM - Posted by Anonymous (not verified)

Just because the average is low doesn't mean the GPUs aren't asking for more power.

If the load is held constant, then the lower power consumption measurements vanish immediately. There’s nothing for GPU Boost to adjust, since the highest possible voltage is needed continuously. Nvidia's stated TDP becomes a distant dream. In fact, if you compare the GeForce GTX 980’s power consumption to an overclocked GeForce GTX Titan Black, there really aren’t any differences between them. This is further evidence supporting our assertion that the new graphics card’s increased efficiency is largely attributable to better load adjustment and matching.

Easy to see if you look through the graphs that its not on occasion its a constant dips and spikes which make up the misleading average. Maxwell just does it better (faster) then Kepler and AMDs PowerTune.

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-980-970-maxwell,3...

October 12, 2014 | 08:44 AM - Posted by Anonymous (not verified)

Shut up, man. If you're gonna so anal, let me be a bit anal too: a 1200W PSU is capable of giving more than 1200W of PSU. Quite a bit more.

A 400W PSU can for example easily power a system that demands 500W, as the 400W is only what is advertised and sold on the product. It can actually go higher.

October 12, 2014 | 08:44 AM - Posted by Anonymous (not verified)

Shut up, man. If you're gonna so anal, let me be a bit anal too: a 1200W PSU is capable of giving more than 1200W of PSU. Quite a bit more.

A 400W PSU can for example easily power a system that demands 500W, as the 400W is only what is advertised and sold on the product. It can actually go higher.

November 14, 2014 | 04:15 AM - Posted by Natalie (not verified)

They are right the figures listed for power supplies are nominal figures.

Between 1990 and 1994 I was stationed in Germany. My barracks room was almost considered a day room for my floor. I had a dual NTSC/PAL 38" TV (CRT type, flat screens came later), a Pioneer laser disk player, 4 VHS systems (1 Pal to NTSC, 1 Super VHS NTSC, 1 Pal/Secam/Mesecam), a super Beta, Fisher AM/FM/turntable/cassette stereo System, and two DBX boxs to route signals between equipment and the television. I also had an Amiga 500 computer. Because of the power cords only the television and the PAL/Secam/Mesecam VCR could be plugged directly into the wall. The rest used NEMA 5-R15 plugs. ALL the rest were plugged into power strips that finally ended up going into a 75W transformer. Almost 24x7 the computer was running, in the evenings friends would be watching television, which meant using at least 2 dbx and 1 or more video recorders. Friday evenings I would always record the German TV shows (soft porn) on the PAL VHS and then while watching something else I rented, I would be converting the German shows from PAL to NTSC format for friends to watch. I would easily be drawing 300-400W through that 75W transformer but it never blew a fuse.

Across the hall a friend would let me iron my uniforms. The irons would be plugged into a 1000W transformer. Check how much power draw for pressing cotton. It is close to 2000W. No problem with the transformer blowing.

While a power supply for a computer might not be as sturdy as those transformers, they are not fragile. They can usually 10-20% more power with no ill effects. Good ones are can take even more especially for short periods of time.

May 20, 2015 | 11:03 AM - Posted by Anonymous (not verified)

I have 4 of these monsters and a 5960X plus two 6tb drives and 64Gb of ram, my Rosewill Hercules 1600W power supply works great. No bottle necks at full load when rendering 4K video.

May 29, 2015 | 01:32 PM - Posted by Anonymous (not verified)

That is ENTIRE SYSTEM power draw. GTX 980's can only draw up to 125% of its 165w TDP which is 205w (unless you modify the video bios.

The GTX 980 is damn efficient and when overclocked to the highest aircooling OC (~1500mhz core, ~4000mhz mem) its as fast as 2 stock clocked GTX 770's or 4 stock clocked GTX 480's.

My 1 overclocked GTX 980 uses 120% TDP (198w) in furmark at 1505mhz core and 3985mhz memory. If you only game with 4 of these, you could use a minimum of an 800w PSU.

While gaming at this OC, I only reach about 173w/105% TDP and 4 way sli is hard to achieve 100% load like 1 or 2 cards can

October 10, 2014 | 05:14 PM - Posted by Anonymous (not verified)

If you had insufficient power supply, wouldn't the system crash rather than throttle?

October 13, 2014 | 03:12 PM - Posted by Sonic4Spuds

Likely yes, though this isn't necessarily the case. It is just likely that the computer isn't engineered in a way to prevent crashing on insufficient power.

October 16, 2014 | 05:48 AM - Posted by arbiter

Lets clear things up the cards that were using 250watts were gigabyte windforce cards, those are not reference cards. Gigabyte is known to loosen up power constrains of the card. the ref card was well under 200watts.

Gigabyte windforce cards come with 2x8pin power connectors were as ref is 2x6pin.

October 10, 2014 | 02:58 PM - Posted by H1tman_Actua1

Anything more then 2way SLI is a total waste for games.

SLI/crossfire in general blows.

October 10, 2014 | 08:04 PM - Posted by Anonymous (not verified)

Well I agree with first comment...

October 11, 2014 | 11:02 AM - Posted by Mike.S (not verified)

From what I have heard & read,Crossfire is still too choppy ,but my experience with 2-way SLI,have been incredible.I agree that 3& 4 way SLI for gaming ,though is not worth the price-performance gains .Strictly for e-peen status.

October 13, 2014 | 11:48 AM - Posted by mesyn191 (not verified)

CF and SLI are about the same in terms of microstutter and game compatability.

They're both a very mixed bag and you shouldn't bother with them unless you really want to game at 4K resolutions, have lots of extra cash lying around, and are willing to put up with a fair amount of driver oddness.

October 13, 2014 | 03:44 PM - Posted by remon (not verified)

Actualy, 290-290X CFX is better than the 7xx series SLI, at least in frame variance.

October 10, 2014 | 03:04 PM - Posted by Prodeous

Wonder if there is a chance for OpenCL/CUDA benchmarks? Like LuxMark/LuxBench, or Blender, or other CUDA/OpenCL?

Just curious if having 4 of these would be useful for this type of workload or if there is some bottleneck for these tasks also?

October 10, 2014 | 05:21 PM - Posted by Anonymous (not verified)

It depends on your code, but most compute task are independent enough that you should see near linear scaling unless there is a communication limitation (not enough pci-e bandwidth between cpu and gpus).

October 25, 2014 | 06:18 AM - Posted by Anonymous (not verified)

Apparently it works quite well:

http://cdn.eteknix.com/wp-content/uploads/2014/09/nvidia_gtx980_SLI_grap...

Source: http://www.eteknix.com/testing-nvidias-geforce-gtx-980-4gb-graphics-card...

October 10, 2014 | 03:08 PM - Posted by tabuburn (not verified)

I've always wondered about why reviewers calculate the scaling percentage of multi-GPU set-ups the way they do.

I noticed that when calculating the scaling for 3-way and 4-way, those values are calculated against the previous value i.e. 3-way results compared to 2-way and 4-way results compared to 3-way. Isn't that basically like saying 3-way performance must double that of 2-way or 4-way performance must double that of 3-way, when adding an additional GPU would just add the performance of a single GPU? It's like asking for magic.

Think about it. 2-way's performance is just the sum total of performance of 2 GPUs minus some driver and/or game code overhead. Wouldn't mean that 3-way and 4-way's performances are the sum total of 3 and 4 GPUs respectively?

Let's use the Crysis 3 test as an example, a single GTX 980 has an average fps of 14.7 at 4K while on 2-way, that is bumped up by 13.5 fps for a total of 28.2 fps with a scaling of 92%. Now, going 3-way gets us up 39.4 fps which means we gained 11.2 fps from adding one more GPU. That 11.2 fps is 76% of the 14.7 fps we got from running a single GPU. Since we only added one GPU going from 2-way to 3-way, wouldn't that mean the scaling of 3-way in this game be 76% instead of 40%? Using the same method to calculate, 4-way is 80% scaling which is even a better result than that of the 3-way one.

October 10, 2014 | 03:23 PM - Posted by Ryan Shrout

This is an interesting idea...let me think on it for a bit.

October 10, 2014 | 03:43 PM - Posted by Anonymous (not verified)

No need to think, that way is just a better way of calculating the fps gain!

October 10, 2014 | 03:51 PM - Posted by Allyn Malventano

We're whipping up some charts to add to the piece - showing ideal vs. measured.

October 10, 2014 | 04:42 PM - Posted by Jesse (not verified)

Yeah, I agree with tabuburn - it looks almost perfectly linear as you add cards in Crysis 3.

October 25, 2014 | 06:29 AM - Posted by Anonymous (not verified)

Why would we calculate it any differently !

Each GPU added would simply increase the number of Cores available, with the 'inefficiency' of them being located on different Cards.

Thus, in a perfect world of scaling, 2 GPUs would be twice as fast as 1 GPU, and 3 GPUs would be 3 times as fast as 1 GPU (etc.).

Why would we think that 3 GPUs should be 2 times as fast as 2 GPUs (etc.), that makes no sense; you would need 4 GPUs to calculate 2 times as fast as 2 GPUs.

This is why Benchmarks that scale well (like LuxMark, FAH etc.) are better than a Game that clearly does not scale well (and thus, it is apparent, that the Programmer was more concerned with simply getting the Code to run on one GPU than having the Code 'behave nicely' with multiple GPUs).

October 10, 2014 | 03:41 PM - Posted by Shortwave (not verified)

That always confused me as well and you explained it well, thanks!

October 10, 2014 | 04:25 PM - Posted by Ryan Shrout

As a response to this comment, I have added a couple of sections to the review on the first and last pages, with some additional graphs to try and demonstrate what is being discussed.

Let me know what you think!

October 10, 2014 | 05:13 PM - Posted by John Mellinger (not verified)

This has nothing to do with the article but Nvidia is going to have a hard time dealing with newer games that are on xbox1/ps4 when they launch given the fact that Nvidia will not have optimized drivers for them.

AMD will have a advantage since the hardware in consoles is X86 arch and is under gcn so this gives AMD a natural optimization over Intel/Nvidia hardware.

Only way Nvidia will have a advantage for now on if it games are true Nvidia The way it's meant to be played games. Shadow of mordor is not a true Nvidia game. It's just published by WB and WB has a partnership with Nvidia so Nvidia got it's name on it just for that reason. The next drivers for SOM, A I and Ryse SoR will be very very interesting and I think you should do a followup article on SOM when the new drivers hit to show performance of green team vs red team again.

October 10, 2014 | 05:30 PM - Posted by Anonymous (not verified)

Giving the performance as a multiple seems like an easy way to understand how good the scaling is. Obviously, a 4-gpu set-up should have a maximum of 4x the performance. Getting ~3.5x is relatively good scaling at the moment. It still doesn't seem to be worth it at all since most games have playable frame rates at 2-way SLI unless you are playing with a ridiculous display set-up.

October 10, 2014 | 05:30 PM - Posted by Anonymous (not verified)

Giving the performance as a multiple seems like an easy way to understand how good the scaling is. Obviously, a 4-gpu set-up should have a maximum of 4x the performance. Getting ~3.5x is relatively good scaling at the moment. It still doesn't seem to be worth it at all since most games have playable frame rates at 2-way SLI unless you are playing with a ridiculous display set-up.

October 10, 2014 | 04:26 PM - Posted by Anonymous (not verified)

This is what I came here to say. But tabuburn beat me to it.

The FPS of the first GPU is the 100%. So let's suppose you have a perfect scaling scenario in an ideal world:
1 GPU = 30 FPS
2 GPU = 60 FPS
3 GPU = 90 FPS
4 GPU = 120 FPS

In this ideal scenario, each of these additional GPUs have 100% scaling because perfect scaling = 100% when counted relative to the FPS of the single GPU.
If you want to count perfect scaling relative to the previous SLI/Crossfire setup, then perfect scaling for 3-way is 50% (3 GPUs / 2 GPUs = 150%) and perfect scaling for 4-way is 33.3% (4/3 = 133.3%).

So for example, when you said Crysis 3 3-way SLI scaling is "only" 39%, it is important to note that if you're calculating it the way you were, the perfect scaling would be 50%, not 100%. So 39% is actually pretty good as far as 3-way SLI scaling goes.

When you say 3-way scaling is "only 39%" and, apparently, thinking that the ideal 3-way scaling under that method of calculation is 100%, then you're expecting the third GPU to have the performance of two GPUs put together, and the fourth GPU to have the performance of three GPUs put together, which doesn't make sense:
1 GPU = 30 FPS
2 GPU = 60 FPS
3 GPU = 120 FPS
4 GPU = 240 FPS

October 10, 2014 | 05:21 PM - Posted by zurv (not verified)

nod nod

October 10, 2014 | 05:28 PM - Posted by Ryan Shrout

Check that update referenced above. Should be clearly now; you are correct.

October 10, 2014 | 06:19 PM - Posted by Anonymous (not verified)

Sweet, thanks.

Also, please note that I think the first method of calculation is more sound, and it gives better graphs with the data points from multiple game benchmarks more clearly displayed as the vertical spread for 3-way and 4-way is also 100% of the graph. That way the data points are spread out over a bigger vertical length and are overlapped less.

Additionally, and perhaps more importantly, it gives an easier and visual way of comparing scaling for a single game between 2-way and 3-way and 4-way setups.

The graph as it is now, we can easily visually compare scaling between two games. But if I want to quickly compare ~visually~ at a glance 2-way and 3-way scaling for a single game, say Crysis 3, then I can't do that.

October 10, 2014 | 06:16 PM - Posted by nathanddrews

I was just thinking the same thing when I read it, that I really don't like the method of measuring scale between each step. It's easier to understand benefit/cost of buying the cards by using basic multipliers. For example, Crysis 3 4K:

1 GPU = 100% = 1x
2 GPU = 192% = 1.9x
3 GPU = 268% = 2.7x
4 GPU = 348% = 3.5x faster for 4X the cost. Benefit/cost ratio of 0.88 overall (1.0 being perfect) doesn't sound so bad, does it?

Again, Sniper 3 4K:
1 GPU = 100% = 1x
2 GPU = 187% = 1.9x
3 GPU = 280% = 2.8x
4 GPU = 360% = 3.6x faster for 4X the cost. Benefit/cost ratio of 0.90 overall (1.0 being perfect). Again, not bad at all!

If anything, using those examples, this tells me that 4-way SLI is worth it.

October 10, 2014 | 06:21 PM - Posted by Anonymous (not verified)

For those 2 games yes but not for the 5 other games.
Battlefield 4
Bioshock Infinite
GRID
Metro: Last Light
Skyrim

October 10, 2014 | 11:52 PM - Posted by nathanddrews

"If anything, using those examples, this tells me that 4-way SLI is worth it."

Why are you restating the obvious?

October 11, 2014 | 01:40 AM - Posted by mesyn191 (not verified)

When doing cost/benefit you have to consider everything and not just cherry pick the most favorable examples. Especially since in the real world PC gamer crowd most no one just plays those 2 games.

October 10, 2014 | 04:30 PM - Posted by Mac (not verified)

Seeing a lot of dropped frames there on the frame times plot, why aren't they reflected in the observed fps chart which should be different to the fraps fps? What am I missing here?

October 10, 2014 | 05:29 PM - Posted by Ryan Shrout

Where exactly do you see dropped frames? You are seeing a lot of frame time variance, but no DROPS that I can see.

October 10, 2014 | 06:05 PM - Posted by Mac (not verified)

Maybe I'm reading it wrong, but when you look at say the bioshock frame times plot when the line goes down to zero multiple times, aren't those dropped frames or runts?Aren't those frames being counted even though they don't show on screen at all?

October 10, 2014 | 06:15 PM - Posted by Mac (not verified)

My mistake, on closer inspection there are slight differences in the observed fps vs fraps fps, I was expecting the difference to be a little more pronounced.plz ignore.

October 30, 2014 | 09:19 PM - Posted by Anonymous (not verified)

"Where exactly do you see dropped frames" statement is proof enough that you, Mr. Shrout, have an inability to understand graphs. You've probably got somebody doing these graphs and interpreting them for you, but he's probably to busy kissing your a$$ then to actually tell you where the frames are dropping. You should fire him.

Second thing I would like to point out. I can't believe there's no Mathematician or Cal Tech Professor in the fields of Engineering, here to tell you that the concept of Percentile versus Frame Variance and Minimum Extreme FPS is a nice idea, but it's wrong. You're basically stating that Percentile is an Input to the function of that curve, and it produces an output in both "Min FPS" or "Frame Variance" graphs. This is why the Chinese are beating the Americans in Math, and you are showing proof of that. This is like building a 5 inch thick metal door that's being displaced by pistons. You're saying that the doors position in the X is independent to it's position in the y, and this ideology is incorrect. Both X and Y in my example here is dependent on Time, and they teach this shit in Dynamics. I'm not sure if you got the idea to do this from FPS being inverted to Frame Time Variance, and said "shet son, lets run with it and see what people think" or you were told this was correct. If you want to be taken seriously, fix this because you won't learn much from it, and it's not correct. I would suggest inverting it. Make Min FPS and Frame Variance the Independent or the input of the graph, and make the Y or Output the Percentile. This way, you are stating that the percentile is dependent on the Frame Variance Data and Minimum FPS. Frame Variance and Min FPS are independent because it's data that you collected from other results at some interval of time. Turn it into a bar graph and not a curve. Say at Domain Intervals (0,1.0FPS;mSxF^-1), we have Y1 amount of percentage. So at 0.0 to 0.1 mSxF^-1, the card appears to either drop frames or produces frames even faster by 0.8%. At 15 FPS in Game "Battlefield 4," this frame rate occurs only 3% of the time while playing this PC Game.

mSxF^-1 = Time (mS) per Frame.

October 30, 2014 | 09:19 PM - Posted by Anonymous (not verified)

Failz

October 10, 2014 | 04:42 PM - Posted by Mustang (not verified)

Why don't you normalise the scaling factors by multiplying the percentage you have calculated by the number of additional cards being used? This would give you a performance scaling as a percentage of the theoretical maximum, a much more useful metric.

For example if you had a linear performance increase and your calculated performance percentages for 2-way, 3-way and 4-way were 80%, 40% and 26.7% respectively then a layman may well believe that the going 3- or 4-way was not worth it, but the graph would show otherwise, normalising as I suggested would yield 80%, 80% and 80% resulting in the benefit being clear to all.

October 10, 2014 | 04:52 PM - Posted by Anonymous (not verified)

2 GTX 980's $1100

3 GTX 970's $990

It sounds like you'd recommend 2 980's even though drivers may improve over time.

October 11, 2014 | 03:40 PM - Posted by QD (not verified)

I'm wondering what there is to recommend. The 970 looks like it uses even more power than the 980- and for what? ANY of these games will run fine on ONE board and you only have a widely accepted scaling benefit with two. Then you have to wonder if the human factors trade off with the power bill.

This is only for a very small percentage of people that have as much ego as money.

October 16, 2014 | 05:40 AM - Posted by arbiter

Less cards in SLI/CF eliminates any issues, As you get 3x SLI/CF the return gets less and less. As can see in some graph's games tend not to be optimized for 3way+ SLI/CF completely. They can do stuff in drivers but better results can be done in game.

October 10, 2014 | 05:15 PM - Posted by zurv (not verified)

Some of your tests are not good for good multi GPUs "scaling" test.
You should really check GPU usage.

metro redux, for example, which does work great for SLI (over non-redux version which was total crap) - it is hard locked at Tri-SLI. I will not use the 4th card at all.

Strange about the bf4. I have Quad gtx 980s (coming from quad titans).BF4 will max out all GPUs. 4 is faster than 3 for me. That said, single player can be wacky sometimes with GPU usage. Might i suggest the practice range (or multiplayer match by yourself).
Without me using Quad i'd not be after to play with ultra and high AA settings (at 4k). BF4 will totally break with some OSD, even to LCDs - but lucky have fancy frame capture stuff :)
(i'm playing a 32" 4k titled screen).

4 cards might not be needed for most.. but min 3 is needed for 4k gaming. It isn't about max fps or even avg.. it is about never dropping below 60. :)

yes, price/scaling for less that 4k is a bit wasteful (haha.. unless compared to people that ran multi Titans :) )

These cards run cool.. use low power and hella OC well. :)
but if you are going 3-4 gtx 980 - go water. :) The noise otherwise is painful.

(without SLI AND 4 cards one would not be able to play Shadow of Mordor maxed at 4k- SLI. The hacked in support isn't great.. but that gets it to 70-80fps. 1-2 card is unplayabe at 4k.

crysis 3 isn't that demanding :) (2-3 cards at 4k is enough)

Don't confuse SLI scaling and how well a game support SLI.
If a game can't max out % usage that is the games problem not a function of SLI scaling. That said, if the game's support for SLI is bad then an investment a multi GPUs isn't a good one.

But monitoring GPU usage % is almost as important as the FPS if you want to talk about SLI scaling.

October 10, 2014 | 05:44 PM - Posted by Vega (not verified)

This test is worthless without positing the CPU speed. A highly overclocked CPU is extremely important for 3-4 way SLI setups.

October 10, 2014 | 11:11 PM - Posted by Ryan Shrout

Stock 3960X processor. Overclocking won't make much difference, sorry.

October 11, 2014 | 01:42 AM - Posted by mesyn191 (not verified)

Yup. Especially at high resolutions like 4K where you're massively GPU bottlenecked.

October 11, 2014 | 01:55 PM - Posted by Vega (not verified)

Completely incorrect assumption. Did you even enable PCI-E 3.0 on the RIVE? 689 watts for a 4-way SLI system, barely 200 watts of power use more over 2-way SLI by adding two more cards.

Test methodology was in error. As for someone who has built and tested some of the fastest 4-Way SLI and crossfire setups in the world, your numbers don't add up.

October 11, 2014 | 02:01 PM - Posted by Vega (not verified)

Completely incorrect assumption. Did you even enable PCI-E 3.0 on the RIVE? 689 watts for a 4-way SLI system, barely 200 watts of power use more over 2-way SLI by adding two more cards.

Test methodology was in error. As for someone who has built and tested some of the fastest 4-Way SLI and crossfire setups in the world, your numbers don't add up.

October 12, 2014 | 01:39 AM - Posted by Anonymous (not verified)

PCI 3.0 doesn't change the results (maybe 5fps). You won't even notice it if it isn't enabled.

October 12, 2014 | 01:39 AM - Posted by Anonymous (not verified)

PCI 3.0 doesn't change the results (maybe 5fps). You won't even notice it if it isn't enabled.

October 11, 2014 | 02:02 PM - Posted by Vega (not verified)

Completely incorrect assumption. Did you even enable PCI-E 3.0 on the RIVE? 689 watts for a 4-way SLI system, barely 200 watts of power use more over 2-way SLI by adding two more cards.

Test methodology was in error. As for someone who has built and tested some of the fastest 4-Way SLI and crossfire setups in the world, your numbers don't add up.

October 10, 2014 | 06:33 PM - Posted by godrilla (not verified)

Nothing new, i have seen a guy use 4 wat sli plus a 5th identical card for physx black titans on evga forums, lmfao. Fyi it works.

October 11, 2014 | 01:44 AM - Posted by mesyn191 (not verified)

Sounds like the PC gamer equivalent of one of those sound system audiophiles who spend $500 on a wooden knob or power cords from the wall to their receiver.

October 10, 2014 | 09:58 PM - Posted by Anonymous (not verified)

I'm curious how the ASUS P9X79 Deluxe can support 4-way SLI ?

If you check ASUS website it clearly states 3-way SLI
http://www.asus.com/us/Motherboards/P9X79_DELUXE/overview/

•3-Way SLI and Quad-GPU CrossFireX Support!

If you download the manual. On page 36 (2-14). It says the 3rd PCI-E is disabled if the 4th is occupied.

October 10, 2014 | 11:12 PM - Posted by Ryan Shrout

We use the Rampage IV Extreme.

October 11, 2014 | 03:24 AM - Posted by Anonymous (not verified)

Isn't this logical scaling?

1 to 2 card scales at 100% because you are doubling the card

2 to 3 cards doesn't mean it will scale an additional 100%, it will be halved to 50% more on an ideal setup. because you're referencing it from a 2nd setup

3 to 4 cards will far scale less because you're comparing it with a 3-card setup. so expect the scaling to go down further 25% on an ideal scenario

maybe
2 to 4 cards will scale between 50% to 100% because you're doubling the performance

October 11, 2014 | 04:18 AM - Posted by raghu78 (not verified)

Ryan
the GTX 980 tri-sli and quad-sli should have been compared with the competition. Tri-fire Radeon R9 295X2 with R9 290X and Quad fire Dual R9 295 X2. Anyway from your previous review its clear that Quad Fire is a beast at 4k when its well supported like in Crysis 3, BF4 and Tombraider. The problem is that game support is not consistent and that is the problem with CF and SLI. Still R9 295X2 with XDMA CF is a fantastic solution for 4K gaming.

http://www.pcper.com/reviews/Graphics-Cards/Radeon-R9-295X2-CrossFire-4K...

October 11, 2014 | 04:19 AM - Posted by Anonymous (not verified)

A big thanks to the readers and also to ryan and allyn for updating the article almost in "real time". It clearly shows the cards are doing a decent job at scaling when properly used. Now the burden is on the game engines to actually handle scaling properly and as we can see from the graphs some of them are not ready yet. I wonder if crossfire behave exactly the same now.

October 11, 2014 | 07:12 AM - Posted by windwalker

Now, that you've done the fun useless comparison, how about some useful ones:
2 x 970 vs 1 x 980
or
3 x 970 vs 2 x 980
What is the fps/$ in those cases?

October 11, 2014 | 01:08 PM - Posted by fkr

now that you have made a uselessly condescending comment, we can make a simple supposition.

they are both the same generation nvidia GPU's so scaling should be the same

google: sli gtx 970

boom

techpowerup article, time to compare crysis 3. yep scaling is the same. 11.2 fps goes to 22 fps

http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_970_SLI/9.html

http://imgur.com/gallery/C5envvd

October 11, 2014 | 01:45 PM - Posted by windwalker

Your comment is the only condescending one.
Get off the Internet and go get laid.

The whole point of visiting sites like this one is so I won't need to assume anything and just see relevant benchmark results.

October 12, 2014 | 01:45 AM - Posted by Anonymous (not verified)

No need to start getting personal here. I think he just misunderstood the 'fun "useless" comparison' part. He took it as if you were negatively critizing the author in a troll-way.

Had to read it twice myself.

October 12, 2014 | 05:51 AM - Posted by Slava (not verified)

I think the scalling is affected by the CPU bottleneck due to stock clocks of the CPU. i would like to see this kind of multi GPU review
performed on overclocked system with good clocks on the CPU and memory.

October 13, 2014 | 05:23 PM - Posted by We (not verified)

I have to love that you guys blame the SLI profile immediately when you're clearly CPU bottlenecked.

Here you have admited it;

"Our power testing was taken from Crysis 3; we obviously had to pick a game that had good scaling up through the 4-Way SLI results in order to make sure each GPU was being utilized as much as possible."

Wait, GPU loads weren't maximized in any other game, lets take Bioshock where was no advantage at all from a 3rd/4th card? That's a CPU bottleneck, end of it. What were the GPU loads? I assume in single mode the GPU was at 99%, since theres a double gain so again 99%/99% in 2Way SLI, since you didn't take any advantage of a 3rd that should be 66%/66%/66% in 3 Way SLI, and 33%/33%/33%/33% in 4Way SLI.

Get each GPU at maximum load, not 30%, let's not blame the drivers when our CPU was holding us back, that's illogical. Tell us the CPU was limiting us, if you still had bad scaling from 2 way sli all the way up to 4 Way SLI and the GPU loads were all max'ed out go ahead blame the drivers.

I'll give an example (1080p ultra settings/780/3930K);

- Single GPU at 50% load with 55 fps: http://i.imgur.com/vJkPwDv.png
- 2 GPU's at 25%/25% load with 55 fps:http://i.imgur.com/HEqn7kv.png

That's not because SLI scales bad, thats because that spot is quite CPU intensive where my CPU bottleneck is extreme, whenever I move somewhere else and remove the FPS cap the GPU's all crank up to 99%/99% load with nearly a lineair FPS scaling between 1 and 2 cards.

A 3960x (no clock mentioned so I assume at stock) isn't an excuse, all CPU's bottleneck to some point and these days its mostly seen in SLI setups.

Leaving an article about SLI scaling a moderator on the geforce forums wrote what could cause SLI scaling and explained pretty much what I've been saying: https://forums.geforce.com/default/topic/532913/sli/geforce-sli-technolo...

October 14, 2014 | 10:44 PM - Posted by Anonymous (not verified)

I've seen some people comment on this already and I am familiar with the concept of diminishing returns, which appears to be the argument about adding a card in SLI, but in your conclusion I still find fault in the statement below:

"Those two games showed expected scaling rates of 90%, 39% and 29% as you go up the multi-GPU ladder. That means for a second investment of $550 (the cost of another GTX 980) you will get

    90% added performance

. Another $550 investment nets you

    another 39%

, while a fourth payment of $550 only gets you

    29% more

."

The problem that I and others seem to have with this statement is that it implies that when you sequentially step up the "GPU ladder", your reference point must change to the ladder step before it. Instead of consistently comparing each additional to the original position, you are changing your comparison each step up.

The first GPU gives you a reference point of 100% performance. The second GPU then gives you 90% of that amount of performance, on top of the first 100%. If you add the third, your calculation would imply you only gain 39% of that original amount on top of that 190%, but in reality, you gain an

    additional 78%

on top of the 90% from the second card, which adds up to 168% improvement over the single card reference point. With the 4th card, then, in this example at least, you have an 87% increase, which puts you at a 255% improvement over a single card, and 355% of the performance.

October 14, 2014 | 10:45 PM - Posted by Anonymous (not verified)

obviously my html did not come out right...

October 14, 2014 | 10:50 PM - Posted by Anonymous (not verified)

Well HTML tags didnt work.

October 16, 2014 | 11:01 PM - Posted by Jimm (not verified)

The default measurement that Ryan and other sites use is the performance gain by going 1 more. Most people are going to want to now the benefit of each additional $550. If a change is small compared to the previous set up in game experience, as a player, that is the meaningful measure.

October 16, 2014 | 11:19 PM - Posted by Paul Jones (not verified)

These are my favorite articles.

Couple of requests:

Could you also include an evaluation of 4K-3D?

There we would expect an ideal of 50% of the frame rate of 4K, what will we actually see?

Also, what about the subjective factor? Yes, I see number going up on a chart, but does that translate into a "wow" factor? Same could be asked about 2K vs 3K vs 4K if the resolution does not translate into a better graphical experience due to textures etc. Numbers can't tell that story.

October 18, 2014 | 09:58 AM - Posted by Jesso2k (not verified)

Hey Ryan, I heard you talking about sli profiles on the podcast. I can play Shadow of Mordor in 4k by going into the Nvidia Control Panel and enabling 'Force Alternate Frame Rendering' in the individual program setting. Running the benchmark after brought the fps up to 70 avg vs 40 fps before.

This is the first time I've tested with a game, I'm wondering if someday you can go back and test that setting with games vs using their sli profile.

Thanks

October 21, 2014 | 06:21 PM - Posted by Anonymous (not verified)

You guys really need to run skyrim with something like Kountervibe ENB or Opthfeldt ENB, a weaksauce card can run vanilla skyrim at good fps. Just a thought anyway.

November 6, 2014 | 01:25 AM - Posted by Matthew (not verified)

Enb has issues with SLI and crossfire last I checked. Otherwise it's something I would like to see as well.

November 12, 2014 | 10:31 PM - Posted by Anonymous (not verified)

Have you considered putting together a Xeon build with dual or quad socket CPUs to alleviate the CPU bottleneck for testing multi-card solutions?

November 14, 2014 | 04:48 AM - Posted by Natalie (not verified)

Have you thought of testing using the NetStor NA255A? That would allow for much cooler temperatures due to much better air flow. Your bandwidth going to/from the processor might be reduced but I think it is still rated at 128Gbps. I know it is an expensive expansion cabinet, but anyone willing to pay upwards to $4000 depending on video card, probably will be willing to pay $2200 for the expansion cabinet if it improves things.

December 16, 2014 | 03:31 PM - Posted by Anonymous (not verified)

Just n2 cool everything with xeon CPU's 64g ram and 4way sli.... be done with it... if one has the money.... multi cards onnly improve the experiance. not by much but can help make it better.... fuck it all

December 21, 2014 | 04:59 AM - Posted by titanroyal99 (not verified)

i would like to see a 3 way SLI nvidia gtx titan black is it possibele

December 26, 2014 | 02:22 AM - Posted by pengobatan tradisional ejakulasi dini (not verified)

an information that is very interesting to read, because it has a great reading the contents of all, it is important for our knowledge sanagt, thanks to share his article.

visit my web : http://robiherbal.com/pengobatan-tradisional-ejakulasi-dini/

December 28, 2014 | 09:58 PM - Posted by obat tradisional gagal jantung kongestif (not verified)

thanks to admin who has been providing the latest information that makes us be increased insight and new knowledge that is wide enough, of course, this makes us become more knowledgeable

http://robiherbal.com/obat-tradisional-gagal-jantung-kongestif/

February 14, 2015 | 01:05 AM - Posted by YOUDIEMOFO (not verified)

I am just glad to see that my quad TITAN's are smashing the hell outta these cards. I have no issues with scaling.....albeit when using 4k that is as the lower resolutions do not even push a quad configuration to use max clocks of the OC'd GPU's. Let alone using "all" of the cards for that matter at lower resolutions.

I am easily getting over 100fps with everything in BF4 maxed, but AA off. And with DSR I get to play at 4k with a higher than average refresh rate as well. 100-120hz 4k anyone....?

February 14, 2015 | 01:11 AM - Posted by YOUDIEMOFO (not verified)

And to me...honestly it seems as if something is wrong with either the system setup, or a software issue. The scaling you guys are getting as opposed to others' is just down right ugly. And those Frame Times....Oh my goodness. I'd have to say that it would be an unpleasant experience there. Something I have no issues with here as well.

http://us.hardware.info/reviews/5623/3/nvidia-geforce-gtx-980-sli--3-way...

BF4 right here.

June 7, 2015 | 07:33 PM - Posted by Chris Katko (not verified)

Sure, I'm late to the party. But it should be noted that whenever one card has data that isn't physically on it's own memory, that memory MUST be shared via PCI Express.

PCI Express is very fast compared to your LAN, or Cable Modem. But compared to a modern video card's memory bandwidth? It's nothing.

So, the more that must be shared, the quicker the PCI Express bus becomes the bottleneck.

BUT, this doesn't have to be the case. A game that takes SLI into consideration can split workloads that don't have hazards (one doesn't depend on data the another must complete first), then minimal memory must be synchronized across the bus. Now, I'm merely speculating--as I have no experience writing AAA games--but if I were to guess, I would say that this is the reason for the large disparity between 0% gain, and 90% of ideal gain for those SLI setups. It all comes down to how those workloads are organized.

So with all of that in mind, it might be useful to compare SLI (even 2X) with differing PCI Express lane speeds.

The motherboard featured has four x16 lanes, and one x8. So try two cards, with one card on 8x (meaning the synchronization will occur at the lowest common denominator of 8x). Also try making the 8x card the "primary" card and see if that changes anything to see if the primary card gets the most data and the secondary gets only supplementary. (I don't know what the memory mirroring/sharing strategy is for SLI.)

Lastly, a switch handles routing PCI Express traffic, and while the motherboard seems like a beast, it's possible that chip is unable to satisfy the throughput of ALL lanes simultaneously. So motherboard chipsets can definitely be the bottleneck in these scenarios.

PCI Express 4.0 is slated for late 2016, and doubles the bandwidth of lanes. It stands to reason a typical PCI Express switch's throughput will increase. So one could, in the future, see if the same cards on a newer 4.0 motherboard ends up with higher throughput. Even though the cards can't run faster than 16x, it's less likely that the switch would be overwhelmed.

This is fun stuff. If you've got the time and hardware, I'd definitely recommend trying these variations. The knowledge could help you in future write-ups.

September 29, 2015 | 12:21 PM - Posted by Anonymous (not verified)

Would Palit GeForce GTX 980 Trio 's be good in SLI? What would be the best configuration (e.g. 2-way sli, 3-way sli)?

November 30, 2015 | 03:09 AM - Posted by Morten Orre (not verified)

For gaming, stick to the fastest single core card you can find (980TI OC'ed on water) to about 1400. If you need high res 120+ FPS use two of these on air or water. Dont go 3 way on SLI, just a big hassle. My setup is 4X Palit 980 TI on water, but not using SLI but for GPU render using Cuda.

November 30, 2015 | 03:04 AM - Posted by Morten Orre (not verified)

I'm currently running 4X Palit 980 TI on water cooling and my scaling is 100% increase pr. card. There is more to multiGPU than SLI and gaming. I'm now using Octane render and this is scaling almost perfectly, even up to 8-12X GPUS. This uses any Cuda core available (even from different cards) and does not use SLI functions.
Not impressed with Nvidias SLI controls, as 3+ GPU setup in SLI often will cause trouble in games as many cant even handle more than one card properly. As a result, you have to physically disable some of the graphics cards on the motherboard, a really akward and not well thought through solution for daily use.

January 4, 2016 | 06:57 AM - Posted by JXI (not verified)

Can you explain the following numbers:

"Looking at just the graphics scores in our results here, on the Fire Strike result, the 2-Way SLI configuration scales by 93%, 3-Way by 35% and 4-Way by 10%. Under the Fire Strike Extreme test we see that move up to 96% for 2-Way SLI, 48% for 3-Way SLI and 30% for 4-Way."

    From my point of view I get:
  • 5937 points 1-way SLI
  • 11628 points 2-way SLI (95.86% increase compared to 1-way)
  • 17160 points 3-way SLI (189.03% increase compared to 1-way)
  • 22315 points 4-way SLI (275.86% increase compared to 1-way)

How do you come to:

4-Way by 10%

February 10, 2017 | 10:35 AM - Posted by NoDebuff (not verified)

I find this really useful because i was considering if to buy another graphics card (sli) and it seems like 2 way has the biggest performance increase than having 2 or more and for the money your going to spend its wise just to get 2 way sli for a performance increase, thx

November 15, 2018 | 06:54 PM - Posted by Very happy 3 way pascal user (not verified)

I'd like to know how many of you that say more than 2 cards is pointless, has actually tried it? If you did try it and didn't use proper settings via 3rd party software for it, well it's no surprise why your experience sucked. 3/4-way is the only way to go if you run 3 screens, Enjoy high graphics and high fps+Hz not 60/60 more like 90/75 minimum.

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

CAPTCHA
This question is for testing whether you are a human visitor and to prevent automated spam submissions.