Original Link: http://www.anandtech.com/show/2728



This marks the final planned installment of our multiGPU exploration. We may (or may not) publish a follow up that looks into CPU scaling across all these parts. What we believe we'll find is that the single GPU solutions will not be anywhere near as significantly impacted as multiGPU solutions which more often hit CPU and other system limitations. We aren't guaranteeing that we'll be publishing the CPU scaling article because we still have some testing to run and this editor is soon to be the father of a second child. We will be working on completing our testing, and whether or not we are able to round this series out with a CPU scaling follow up, we will definitely be exploring CPU scaling further in future articles.

It is important that we remember, for now, that much of the diminishing return on what 3 and 4 GPU systems can deliver comes in the form of system limited performance. With single GPU systems, we expect that there is a wide range of CPUs we can select that will deliver nearly the same performance. Putting less money into the CPU than the GPU makes a lot of sense for gamers who don't need the CPU power for other tasks. But does the same hold for multiGPU systems? Maybe and maybe not. We do know that with the highest powered CPU we can buy we certainly have a good number of system limited situations.

One of the things that people who invest in the highest end multiGPU systems get is more longevity. Gamers with 3-way GTX 285 hardware will be able to go quite a while without upgrading. Of course, this has to be balanced with advancements in technology. Will 3x GTX 285 still be worth it after we have DX11 hardware and games and there are better graphics options for newer models? Additionally, gamers who want and employ this type of solution are highly likely to upgrade as often as possible to the highest end hardware possible, so the longevity issue might not be as relevant as it is with more affordable multiGPU solutions.

On the plus side, if gamers who must have the bleeding edge buy the highest end equipment they only need to make a significant investment the first time. Imagine a gamer bought 3x GTX 285 parts when they came out in January. Within a year we expect new models to come out that will be able to best the GTX 285 in performance, but the GTX 285 hardware will still be worth a significant amount. These used cards can be sold and the profits can be reinvested in new graphics hardware. This isn't as easy or useful with lower end graphics cards as you don't get the same return on your investment.

I'm not saying you'll make money or break even this way, but if you aggregate the cost over the long term, you'll be spending less on average per purchase than if you just make one high end purchase and then run it until it's worthless before you buy again. Whether or not you'll spend more or less on the whole would require further analysis though.

Now, we advocate value quite a bit on our site. People who want the absolute highest end can easily look at the graphs and simply know what they want. We don't really see much "value" in these high end parts beyond being as fast as possible. It's cheaper to buy efficient lower end hardware that gets hugely playable performance in all the games we tested. A good high end 2-way solution can go a long way while the diminishing returns of 3 and especially 4-way systems make the cost of the slightly higher performance is more than we can recommend in good conscience. But there's a market for it, so we'll take a look at it.

For the majority of our readers, though, this article will reinforce the fact that, while the highest possible end might be nice to dream about, it's really tough to justify the cost, especially when most people are on some sort of budget. So let's get a glimpse of that dream.



What We Couldn't Cover

Our tests will include the GeForce GTX 295 Quad SLI, GeForce 9800 GX2 Quad SLI, Radeon HD 4870 1GB Quad CrossFireX, and Radeon HD 4850 Quad CrossFireX. We were unfortunately not able to test the 4850 Quad with 1GB per GPU because we didn't have 2 of the 4850 X2 2GB cards. This would undoubtedly have made the 4850 look a little stronger in Quad at 2560x1600 (where it really counts). While it wouldn't compete for the highest end performance, the higher memory Quad 4850 is certainly of interest to us after seeing the value in two of them. But we really don't expect any Quad option to deliver on bang for buck metrics.

Past that, we also didn't include Race Driver GRID this time around. Due to our continuing issue with FRAPS, we couldn't record performance data for either of our Quad NVIDIA solutions. We didn't feel that presenting the data from the game with just AMD hardware was highly useful, but it is worth mentioning that just looking at the numbers we could tell that the Quad NVIDIA solutions performed slower than the AMD solutions. I do apologize for a lack of quantitative data, but sometimes that's how it goes. We will continue to try and collect this data and we may do something with it down the line if we are successful.

Our first article explored 1 to 2 GPUs. The second looked at 1 to 3 and 2 to 3 GPUs. This one only focuses on the performance improvement from 2 to 4 GPUs. The reason for this is that there are no single GPU versions that exactly match half a GTX 295 or half a 9800 GX2. We can see when things don't scale and how they scale differently from 3 way by looking back at the previous article if people want, so having a lopsided analysis that included some metrics for AMD and not NVIDIA didn't seem quite right.

Just like the diminished returns we saw when moving from 2-way to 3-way, we see more diminishing returns when moving from 3-way to 4-way. The way we can get a feel for that more directly is that we see much less scaling when moving from 2-way to 4-way than when moving from a single GPU to two (even though the theoretical performance improvement is the same).

This time around, we didn't zero the value data when performance didn't meet a threshold. We know some people liked that way of doing it, but value really isn't a focus of a 4-way GPU shootout anyway, so we feel that the data is more academic on its face. This article rounds out our data and has all the performance numbers for all the parts we've looked at, while our analysis focuses on 4-way. We are still actively refining our approach to representing value moving forward, so your feedback is not only welcome, it is greatly appreciated.



Who Scales ... And Timing

In previous articles, we took a look at some stats on how many tests scale more or less than a certain threshold. Well, it gets a little trickier here to make sense of everything, so rather than pick points ourselves, we decided to list a bunch of them. In the chart below, we've listed the number of tests that fail to scale better than the percentage listed at the top of the column. Lots of tests fail to scale at what we would call a reasonable percentage, but people looking at these parts have a different definition of reasonable.

The maximum scaling percent is 100% just like with scaling from 1 to 2 GPU. But fewer games scale past 2 GPUs, and of those that do, fewer scale as near linearly past 2 GPUs. And to top that off, many tests that do scale at all scale right into a system limitation. A good chunk of games fail to scale past 5%, and fully 13 out of 18 tests fail to scale beyond 50% in each case we tested.

<2.5 <5 <10 <15 <20 <25 <33.3 <50
NVIDIA GeForce GTX 295 Quad SLI 5 5 6 7 8 9 10 13
NVIDIA GeForce 9800 GX2 Quad SLI 7 7 8 8 8 8 8 13
ATI Radeon HD 4870 1GB Quad CrossFire 5 5 5 7 9 11 12 13
ATI Radeon HD 4850 Quad CrossFire 6 7 9 10 10 10 12 13

Looking at the lower end we can see a bunch of tests fail to scale at all. Also, at 33%, many fewer situations scale at this rate than when moving from 1 to 2 GPUs. Clearly 4-way multiGPU solutions are not designed with anything but maximum performance in mind. Scaling isn't as important as the fact that these solutions can provide some degree of higher performance in some situations.

We would also like to note that when paying ridiculous amounts of money for not quite as ridiculous performance gains, the robustness of the solution is of very high importance. No one wants to pay over $1000 and get a solution that sometimes provides good scaling and sometimes degrades performance. Neither AMD nor NVIDIA are immune to this, but we would like to see this issue tackled in more earnest beyond simply noting that SLI and CrossFire can be disabled if trouble arises.

NVIDIA does have an advantage at this level though. We would love to see AMD get their driver act together and consistently have drivers that provide good scaling and performance in newly release AAA titles on launch day. We would also love to see them refine their driver development model in order to make sure that improvements released as hotfixes always make it into the very next WHQL driver released (which is currently not the case). Everywhere else, this is merely a slight annoyance that people may take or leave. At the highest of the high end, however, a delay in getting good scaling or the need to use less recent drivers that contain more recent fixes (and juggling which is which) can prove more than just a trifle. For such a high price, NVIDIA delivers a better experience on this count.

Additionally, until OpenCL matures, CUDA is a better GPU computing alternative to what AMD offers, and PhysX can provide additional flexibility now that more titles are beginning to adopt it. Actually, this is the space in which we currently see the most value in CUDA and PhysX, as those in the market for equipment this high end will be more interested in these niche features that don't have broad enough support or large enough current impact for us to heartily recommend them as a must have for everyone.

Technophiles (like myself) that are willing to put this kind of money into hardware often get excited about the hardware on a more than practical level. The technology itself, rather than the experience it delivers, can often be a source of enjoyment for the end user. I know I like playing with PhysX and CUDA in spite of the fact that these technologies still need broader support to compel the average gamer.

Performance, itself, cannot be ignored, and is indeed of the highest importance when it comes to the highest end configurations. We will include the value graphs, but we expect that the line closest to the top of the performance charts are the key factor in decision making when it comes to Quad GPU options. The troubles that come with maintaining a 4 GPU configuration are not worth it if the system doesn't provide a consistently top of the line experience.



Prices, Stutter and The Test

We've added our four new test cases to the price lineup. Naturally they are not cheap.

Cost of Graphics Solution

As we mentioned in the 3-way article, our experience with stutter increased with the number of cards. Two way seems to be the smoothest of the multiGPU options, and we ran into the most problems with 4-way. Both AMD and NVIDIA showed some stuttering in Crysis, and there were issues in other games where scaling didn't happen the way we would have liked. Honestly, gamers who choose 4-way options will need to be hands on to get the best experience, disabling SLI and CrossFire when they get in the way of themselves.

Thus the price to pay for these solutions is not just higher in terms of money, but higher in terms of the effort needed to maintain a positive experience. NVIDIA gives us options to entertain ourselves by using some of our hardware for PhysX rather than SLI if SLI doesn't happen to work out as well as expected. This is definitely a plus at the very high end, as simply disabling hardware completely is dissatisfying in light of the cost.

Our test system is the same as it has been for the previous articles.

Test Setup
CPU Intel Core i7-965 3.2GHz
Motherboard ASUS Rampage II Extreme X58
Video Cards ATI Radeon HD 4870 X2
Sapphire ATI Radeon HD 4850 X2 2GB
ATI Radeon HD 4870 512MB CrossFire
ATI Radeon HD 4850 CrossFire
ATI Radeon HD 4870 1GB
ATI Radeon HD 4870 512MB
ATI Radeon HD 4850
NVIDIA GeForce GTX 295
NVIDIA GeForce GTX 285 SLI
NVIDIA GeForce GTX 280 SLI
NVIDIA GeForce GTX 260 SLI
NVIDIA GeForce 9800 GTX+ SLI
NVIDIA GeForce 9800 GX2
NVIDIA GeForce GTX 285
NVIDIA GeForce GTX 280
NVIDIA GeForce GTX 260 core 216
NVIDIA GeForce GTX 260
NVIDIA GeForce 9800 GTX+
Video Drivers Catalyst 8.12 hotfix
ForceWare 181.22
Hard Drive Intel X25-M 80GB SSD
RAM 6 x 1GB DDR3-1066 7-7-7-20
Operating System Windows Vista Ultimate 64-bit SP1
PSU PC Power & Cooling Turbo Cool 1200W

And now on with performance.



Age of Conan Analysis




1680x1050 1920x1200 2560x1600



1680x1050 1920x1200 2560x1600



1680x1050 1920x1200 2560x1600


Call of Duty World at War Analysis



1680x1050 1920x1200 2560x1600



1680x1050 1920x1200 2560x1600



1680x1050 1920x1200 2560x1600



Crysis Analysis




1680x1050    1920x1200    2560x1600





1680x1050    1920x1200    2560x1600





1680x1050    1920x1200    2560x1600




Fallout 3 Analysis




1680x1050    1920x1200    2560x1600





1680x1050    1920x1200    2560x1600





1680x1050    1920x1200    2560x1600




FarCry 2 Analysis




1680x1050    1920x1200    2560x1600





1680x1050    1920x1200    2560x1600





1680x1050    1920x1200    2560x1600




Left 4 Dead Analysis




1680x1050    1920x1200    2560x1600





1680x1050    1920x1200    2560x1600





1680x1050    1920x1200    2560x1600




Power Consumption

As expected, 3 and 4 way options draw tons of power even when idle.

Idle Power

Quad 4870 1GB CrossFire posts an amazingly high 721.8W of power drawn at the system level under 3dmark Vantage using the POM shader test. Keep in mind, again, that this test does not stress the CPU or memory in significant ways like a game would. Expect 50W to 100W of additional power drawn when the card is stressed using a game. This means that a 750W PSU can't even pretend to be enough for 4-way 4870 1GB. When running 3 and 4-way solutions, we strongly urge those interested in this kind of a setup to invest in a PSU that can deliver more than 1KW.

Load Power



Final Words

So this may come as a surprise to some, but the AMD Radeon HD 4870 1GB Quad CrossFire leads in our benchmarks when focusing on the resolution that matters for this hardware (2560x1600).

While driver issues and a lack of other "stuff" like PhysX and CUDA to do matter with GPUs in situations where hardware doesn't scale, the AMD solution leads the GeForce GTX 295 in more benchmarks (Age of Conan, Left 4 Dead, and Far Cry 2), and ties the NVIDIA solution in one title (Fallout 3). Not shown in our numbers is Race Driver GRID, as we have a continuing issue in FRAPS that gets in the way of recording performance numbers with 4-way NVIDIA solutions. We were able to watch frame rate, however, and it was clear that the NVIDIA hardware didn't reach the performance levels of AMD hardware in GRID.

Certainly this isn't a sweeping victory for AMD, and the outcome, because it is close, rests incredibly heavy on the benchmarks we chose and were able to run. Different titles may have produced different results. Thus there is no clear winner in terms of absolute performance. This will depend greatly on title preference. It is worth noting, however, that when Quad GTX 295 leads Quad 4870 1GB, the NVIDIA card comes in at the very top in terms of performance more often than does AMD. But the dark horse in the 4-way focused article is the 3-way high NVIDIA GPUs.

The 3-way GTX 280/285 leads the 4-way GTX 295 in half our tests: it's a wash and it's either slightly cheaper or slightly more expensive depending on the specific flavor. The 4-way Radeon HD 4870 1GB only leads 3-way GTX 280/285 in 2 out of the 6 tests, though it ties in one of them (Fallout 3 again). If GRID were added back in, it's likely the playing field would be completely even on that count.

If you want an added twist, moving from 2-way to 4-way, AMD tends to scale better at 2560x1600 than NVIDIA. Whether that's because of lower baseline performance of the 2-way option and less system limitation at the high end, it's still impressive that the playing field is this even.

So what's the bottom line? Wow ... It's very hard to say that the differentiator is only performance itself. But as we had less trouble with 3-way than 4-way, so our very slight preference for this one is the 3-way GeForce GTX 285. Overclocked hardware will get you even further into the stratosphere. Enjoy the ride.

If you don't happen to have a motherboard that supports 3x double-slot x16 physical PCIe cards, 4-way will have to be the option. In that case, Quad HD 4870 1GB scores points for keeping up with the Joneses, scaling, and bang for the buck. In terms of performance per dollar, which some people may not care about at these top end price points, AMD leads. At the same time, we must consider that heavy investors like things to play with and PhysX and CUDA do add a potential benefit over AMD that some enthusiasts may like.

So who's got the true halo? Who can provide the best highest-possible-end option? In spite of our leanings and recommendations and considerations, It's a wash. This one goes down in the history books as a battle for the high end that will come down to brand preference.

Log in

Don't have an account? Sign up now