AMD CrossFire Vs. Nvidia SLI Scaling Analysis

Name: Test System And Benchmarks - AMD CrossFire Vs. Nvidia SLI Scaling Analysis
Author: Don Woligroski

by Don Woligroski October 18, 2010 at 1:00 AM

Test System And Benchmarks

Our charts list the graphics cards based on their retail prices, from the most expensive configuration on the top to the cheapest on the bottom. Our testing involves a number of graphics cards from many manufacturers, so we are setting all of the clock speeds to reference levels to keep things on an even playing field.

There is one testing configuration that we have to simulate: two Radeon HD 5850 cards in CrossFire. We only received a single Radeon HD 5850 in time for testing, so we’re pairing the card with a Radeon HD 5870. AMD has confirmed that the superior card will be crippled when these models are paired in CrossFire, resulting in Radeon HD 5850-class specifications, using 1440 of its 1600 stream processors and 72 of its 80 texture units.

We've set the clocks of both cards to Radeon HD 5850 reference speeds; the resulting performance should be exactly the same as two of these cards in CrossFire. The results show the performance exactly where we would expect a Radeon HD 5850 CrossFire configuration to be relative to the Radeon HD 5830 and Radeon HD 5870.

			Test System
CPU			Intel Core i7-920 (Nehalem), 2.67 GHz, QPI-4200, 8 MB Cache Overclocked to 3.61 GHz @ 172 MHz BCLK
Motherboard			Gigabyte X58A-UD3R Intel X58, BIOS version FA
Networking			Onboard Gigabit LAN controller
Memory			Mushkin PC3-10700H 3x 2048 MB, DDR3-1376, CL 10-10-10-25-1T
Graphics			Nvidia GeForce GTX 480 Reference 1536 MB GDDR5 Zotac GeForce GTX 480 AMP! Edition 1536 MB GDDR5 HIS HD 5870 iCooler V Turbo X Radeon HD 5870, 1 GB GDDR5 Gigabyte GV-R587SO-1GD Radeon HD 5870, 1 GB GDDR5 MSI GeForce GTX 470 1280 MB GDDR5 Gigabyte GV-N470SO-13I GeForce GTX 470, 1280 MB GDDR5 Gigabyte GV-R585OC-1GD Radeon HD 5850, 1 GB GDDR5 AMD Radeon HD 5830 Reference 1 GB GDDR5 2x Gigabyte GV-N460OC-1GI GeForce GTX 460, 1 GB GDDR5 (all clock rates have been set to reference specifications for the purpose of benchmarking)
Hard Drive			Western Digital Caviar WD50 00AAJS-00YFA, 500 GB, 7200 RPM, 8 MB cache, SATA 3.0Gb/s
Power			Thermaltake Toughpower 1200 W 1200 W, ATX 12V 2.2, EPS 12v 2.91
Software and Drivers
Operating System			Microsoft Windows 7 Ultimate 64-bit
DirectX version			DirectX 11
Graphics Drivers			AMD Catalyst 10.7 Beta, Nvidia GeForce Driver 258.96 WHQL

Previous Next

Summary

About the author

Don Woligroski

This thread is closed for comments

106 comments

Comment from the forums

Your comment

reprotected

So... Nvidia won?
joytech22

Now i can safely say that i think i made the right choice
(GTX470 SLI)
jrharbort

Nvidia wins this war thanks to the larger standard framebuffer. ATI needs to pick up in this area a bit, as 1GB framebuffers have been around for a while now. Looking foward to another article based on the next gen cards a few months from now. Should be interesting. =)

Good job to both ATI and Nvidia on this generation.
gordo_46

Quote:

BUT.. i would have preferred 2-3 way SLI with GTX460's.

gtx 460 can only do 2 way sli
gkay09

The SLI scaling is one place where Nvidia wins almost all the time, but ATI has improved a lot though, it still need to work harder if they want to have a complete win over Nvidia...
joytech22

gordo_46gtx 460 can only do 2 way sli

Aww damn that's right..
gordo_46

yeah it a shame that it can only do 2 way sli. We would get 3 way gtx 470 sli level performance at amd radeon power consumption
Nice article. I've made some charts, based on the data in this article, that reflect SLI and CrossFireX scalings. Nvidia's victory is clear cut at 2560 x 1600.

1920x1200
http://www.imagebam.com/image/741eca102680169

2560x1600
http://www.imagebam.com/image/0295fe102680172

Also, when it comes to GPU discussions, the comment sections of Tom's and Fudzilla are swarmed by fanboys of 1 certain team. (Can you tell which? It's already obvious here, but much more so at Fudzilla.) Competition is good. Don't downrank comments just because they favor the "other" team.
Twoboxer

12 million people with WoW subscriptions, new expansion due 12/7/10, lot's of vid cards being bought for that purpose . . . I know its not easy, the format may have to be different, but can you guys consider including WoW in the benching runs?
Yet another reason AMD would be happy to never have board of theirs allow SLI
cangelini

Twoboxer12 million people with WoW subscriptions, new expansion due 12/7/10, lot's of vid cards being bought for that purpose . . . I know its not easy, the format may have to be different, but can you guys consider including WoW in the benching runs?

I'd like to hear some discussion on the best way to reliably test WoW, actually.
Just a quick question... in the 2560x1680 chart, how can the gtx 480 have a minimum higher than the average?
joytech22

holto243Just a quick question... in the 2560x1680 chart, how can the gtx 480 have a minimum higher than the average?

Divide by zero O.o
lol wow, if you are still playing that game, then blizzard has you by the balls.
tom's hardware also said that rampage II extreme was better than gigabyte x58-ud5.. And I also ask you... how many of you have monitors that support 2560*1680 resolution . Try to be objective!
super_tycoon

angel1980Nice article. I've made some charts, based on the data in this article, that reflect SLI and CrossFireX scalings.

first, i'd like to thank angel1980^^ for doing what tom's should have done... seeing as how scaling itself was implied to be the thing tested from the article title.

Annoyance aside, I also crunched the numbers, but was left a dilemma as to whom to declare the winner of the multi-gpu scaling. In the end, I wanted similar performance because higher performance = more bandwidth that must be managed and I wanted the same amount of vram. This conveniently leaves the gtx 460 and two amd cards on either side. (But makes my whole analysis somewhat pointless since there's so little data)

Pushing on, using my awesome-o math skills, I derived the following results:

Using the 1080p res, the GTX 460 1GB had scaling values of 1.590 and 1.473, avg, min, respectively. The 5830 and 5850 averaged had values of 1.572 and 1.488.

Using the big 2560x1600 resolution, the GTX 460 1GB had values of 1.768 (!!) and 1.414. The lower 58 had 1.454 and 1.216.

My conclusions? (Completely biased)

At what I think most gamers play at and are comfortable and familiar with, 1080p, AMD's greatest and Nvidia's greatest (the GTX 460 did scale better than the 470 and 480 consistently) architectures scale equally well. However, looking at the 2560x1600 res, we see that Nvidia takes a huge lead... which I suspect indicates Nvidia's SLI tech is better at handling large amounts of data while AMD's is probably a little quicker. (Just a hunch really, you hopefully noticed that this is completely biased, I am an AMD fan) (But the scaling on the 460 was very impressive, hence !!)

I am not surprised at all, it was oft muttered that SLI was superior to AMD's CF. However, what I found surprising is how little that really mattered. The games that favored architectures still favored the representative scaling technology. Simply, look at piggy bank and adjust your reality accordingly. AND, I haven't seen any numbers, but I know AMD released a catalyst update to boost SC2 CF performance, so feel confident in pretending AMD's numbers are truly a smidgen higher.

My shameless endorsement of AMD's overall tech though is simple; in a multi-card situation, it uses a whole lot less power and won't inspire you to break out eggs and bacon. And I must confess I barely sleep thinking of what golden deity AMD will let me waste my money on in just a few days.....

TL;DR Scaling still sucks. At lower resolutions, it's pretty much the same unless you get cards with extra ram, then I don't know, but perf++ regardless. At high bandwidth things, like high res or high performance, sniff, Nvidia takes the cake.

Last bit to holto243:

holto243Just a quick question... in the 2560x1680 chart, how can the gtx 480 have a minimum higher than the average?

Those numbers are relative to the performance of the pretty weak 5830. It simply means the GTX 480 is simply better (ignoring raw power) at handling min framerates in the games tested.
bin1127

Can they make a cheaper SLI/Crossfire edition card that has no ram acting only as a supplemental for the main one? or sell them in pairs so one has 2 gb and the other none since that's how the pairing works.
feeddagoat

Any way to rebench the games where ATI stuttered in the opening seconds, just to get an idea of what their actual min frames are? Would it also be possible to bench using surround/eyefinity? after all, those who get and SLI/crossfire solution will most likely want to do that.
hari_41

both companies doing well
ares1214

Nvidia did well, but thats to be expected, whats more impressive is how much better ATi has gotten! They use to be mediocre with this, and i was really hoping they would overome it with HD6OOO, but it looks like they dont have to improve much.
alyoshka

I really can't say anything, from the price point of view and the current point of view I still feel the ATI wins....
GeoMan

Slightly misleading article name Tom‘s, this is absolute SLI and CrossFire performance, not scaling. SLI and CrossFire scaling to me is the % improvement between running a single card and the same card in SLI or CF.

For anybody who’s interested in those numbers I did some quick and dirty number crunches on the 1200*1080 figures and ignored Far Cry 2 and Star Craft 2 because they’re CPU limited and thus have no influence on CF or SLI.

Card min% max% avg%
480 78 66 71
5870 78 61 68
470 64 59 61
5850 80 57 67
460 62 56 59
5830 70 58 63

Overall I’d place more importance on the improvement to minimum frame rates because with these cards this will be the greater influence to gaming experience. To me the real winners here are the 5850, 480 and 5870 with an honourable mention to the 5830 for an improvement in min frame rate. With the glaring exception of the 480, AMD seems to have better scaling cards and taking price into consideration better bang for buck..

Sorry about the horrible looking table, but it just doesn’t copy paste well and tabs don’t work
eyefinity

Why do this 5 days before the new 6 series is released?

Will you be updating the results then?
Chris_TC

At 2560x1600 the GTX 460 SLI has better minimum frames and only 5% worse average than the 5870 Crossfire.