NVIDIA Claims CrossFire Still Broken

[Apoptosis]

A friend over at NVIDIA sent me this e-mail... I figured copy/pasting this for our readers is only fair and I'm sure NVIDIA sent this out to key press to let them know of their thoughts. I feel it is only fair to pass information like this directly to the LR community, so they can see the marketing direction NVIDIA is going and what they are talking about.

I usually do not forward stories around, but this one was very enlightening. NVIDIA has long said that multi-GPU solutions are only as good as their driver support. That was proven in the Multi GPU World Tour in 2006: http://www.slizone.com/object/slizone_m ... press.html. Seems little has changed in a year and a half.

As a part of its monthly 3D digest iXBT published a very comprehensive analysis of 3870X2 (and CF in general) in 50 games and more than 100 game scenes.
They checked both rendering quality and performance
For each game scene they ran 3870X2 with one GPU (Cat A.I. off) and two GPUs (Cat A.I. on) and compared to 8800 Ultra results
For all games they checked two or more scenes in order to be more accurate with conclusions (scenes where carefully selected not to be CPU bound).
You can check results for each game in the article

Some conclusions from the article:

Comparison with GeForce 8800 Ultra
· GeForce 8800 Ultra was faster in 31 out of 49 (63.3%) games
· Cards showed roughly equal performance in 11 games (22.4%)
· 3870X2 was faster in only 7 out of 102 (14.3%) games
· In all DX10 games GeForce 8800 Ultra was faster

Out of all 102 game scenes enabling CrossFire on 3870X2 delivers
- performance drop in 23 scenes (22.5%)
- no performance change in 44 scenes (43.2%)
- performance boost in only 35 scenes (34.3%)

Games with no performance change with Catalyst A.I. / Crossfire enabled (36%):

1. BioShock v1.1, DX10
2. Command & Conquer 3: Tiberium Wars v1.09
3. Call of Juarez v1.1.1.0, DX10
4. Colin McRae DIRT v1.2
5. Dark Messiah of Might and Magic v1.02
6. Ghost Recon Advanced Warfighter 2 v1.05
7. Gothic 3 v1.6
8. Hellgate: London v1.35, DX10, 64-bit
9. Need for Speed: ProStreet v1.1
10. Neverwinter Nights v1.68
11. Prince of Persia: The Two Thrones
12. Sin Episodes: Emergence
13. Splinter Cell v1.3
14. Splinter Cell: Double Agent v1.02
15. Star Wars: Jedi Knight - Jedi Academy v1.01
16. Star Wars: Knights of The Old Republic v1.03
17. Star Wars: Knights of The Old Republic 2 - The Sith Lords v1.02
18. Unreal Tournament 3 v1.1

Games with significant performance drop with Catalyst A.I. / Crossfire enabled (18%)

1. Company of Heroes: Opposing Fronts v2.201, DX10
2. Condemned: Criminal Origins
3. The Elder Scrolls 4: Oblivion v1.2.416
4. Gears of War v1.1, DX10
5. Kane and Lynch: Dead Men
6. Splinter Cell: Pandora Tomorrow v1.31
7. Test Drive Unlimited v1.66A
8. Thief: Deadly Shadows v1.1
9. World in Conflict v1.0006, DX10

Unstable in games –showed both performance boost and significant performance drop depending on the scene with Catalyst A.I. / Crossfire enabled (12%):

1. Call of Duty 4: Modern Warfare v1.4
2. Half-Life 2: Episode Two
3. Hitman: Blood Money, v1.2
4. Jericho
5. Lost Planet: Extreme Conditions v1.004, DX10
6. Splinter Cell: Chaos Theory v1.05

Looking at only DX10 games they found
- only one game (Crysis) showed some performance increase
- one game (Lost Planet) showed mixed reuslts
- three games (BioShock, Call of Juarez, Hellgate: London) didn't show any perf changes
- three games (37.5%) showed performance drop (Company of Heroes: Opposing Fronts, Gears of War, World in Conflict)

The complete story is here

http://www.ixbt.com/video/itogi-video/t ... 70x2.shtml

What are your thoughts?

[Bwall]

I don't know where to start except to say this review had to be sponsored by Nvidia like the World multi-GPU tour.

1.) Test system is extremely out of date. CPU performance will hold back SLI/CrossFire due to the driver overhead and feeding both systems. Do we even know if that board is stable with a 3870X2? Using Nvidia's own 7950GX2 meant you had to do a BIOS update on the DFI nForce4 boards.

2.) I think every review I read shows increases in BioShock. BioShack can't be timedemo'd so I have to wonder about the rest of their results.

3.) Resolution.. come on, 1600x1200? Again, CPU limitation in most games with cards this powerful.

4.) The last time I spoke to ATI regarding CAT AI, turning it to ON forces AFR. Not all games will run with AFR or (run poorly) and could be a leading reason why they had stability issues.

5.) Games that seem to have a cap in FPS like Command and Conquer 3 were included in their "shows no performance increase" and "were the same speed as the Ultra." I'm not sure why NFS Pro-Street was capped at 30FPS as I have not experienced this.

6.) The biggest point that gets me and what smells the funniest...where are the SLI results?

[DX]

It is a classic fear and uncertainty tactic that is very under handed. What I find most telling is that they don't state the test methodology used, the test system used and well no real details. Nor do they explain the obvious difference in benchmarks between theirs and the tests that the community have already conducted. I think a scathing reply should be made. That sort of behavior from a large powerful firm should be well below them. This now after failing to provide proper drivers for Vista in a timely manner just so that they can push the new products tells me that not only do they not respect the customers that they have in anyway that they also think that we are brainless idiots just drooling for the latest card just because it has the Nvidia stamp on it. As another reviewer stated in summary (name escapes me at this time) ATI had drivers ready and working for present and past products to work with Vista. Where as Nvidia and Creative didn't and to some extent still don't for past products. I tell you what, I look for is performance and if ATI is beating the pants off Nvidia again I'll go that route for sure. I still might just to state to Nvidia that I don't condone what they are doing by starting a campaign of FUD.

Honestly, isn't this the same kind of crap that 3DXF did shortly before they ended up tanking? Remember how they started to just rely on marketing crap instead of tech?

[ibleet]

I have my thoughts, but what do you make of this Nate? I know you can be objective.

[Zertz]

I heard AMD claims Phenom is faster than Intel's quad core too! ;)

No but seriously, like Bwall said, their test system is in no way appropriate to benchmakr such high-end hardware. I bet SLI wouldn't show any better improvements in any of those games...

Also, some games are just poorly optimized. Hellgate London is a good exemple, both SLI and CF show no improvement at all and DX10 performance is laughable no matter what card you use.

[Apoptosis]

I agree with Brian about the test system being a bit old as an AMD Athlon 64 X2 4600+ is sure to be the bottle neck and as a result are any of the results valid? I think it is safe to say that anyone that buys a $600 GeForce 8800 Ultra or a $500 Radeon 3870 better not be using a $20 procssor. A ton of improvements have been done to CATALYST drivers since 8.1, so it's hard to come to a conclusion when 15 or so games have performance improvements and/or bug fixes since the article was written. Just for fun I rounded up all the major enhancements in CATALYST 8.2 and 8.3 drivers.

Game fixes and improvements alone in CATALYST 8.1 and 8.2 drivers are as follows:

Performance Improvements:
* Call of Juarez: Performance scores increased up to 20% on ATI CrossFireX™ configured systems containing an ATI Radeon™ HD38x0 series of product
* Company of Heroes (DirectX 10): Performance scores increased up to 17% (especially at lower resolutions), on ATI Radeon™ HD38x0 products
*Company of Heroes: DirectX 10 performance improves as much as 20% on the ATI Radeon HD 2000 Series and ATI Radeon HD 3000 Series of products. DirectX 10 performance improves as much as 30% on CrossFire configurations using the ATI Radeon HD 2000 Series or ATI Radeon HD 3000 Series of products
* Crysis (DirectX 10): Performance scores increased up to 15% on all supported ATI Radeon™ products
* Lost Planet (DirectX 9): Performance scores increased up to 36% across all supported ATI Radeon™ products and in ATI CrossFireX™ configured systems
* Shadermark 2.1: Performance scores increased up to 35% across all supported Radeon™ products and in ATI CrossFireX™ configured systems
* Unreal Tournament 2004: Performance scores increased up to 10% on ATI CrossFireX™ configured system containing an ATI Radeon™ HD2400, HD2600, or HD3400 series of products

Bug Fixes:
* Elder Scrolls IV: Oblivion: Playing the game on a system running Windows Vista with default game settings no longer results in the gadget sidebar bleeding through.
* Enemy Territory: Quake Wars: Playing the game under Windows Vista (64 bit version) no longer results in corruption being noticed when loading multiple maps.
* Enemy Territory: Quake Wars: The game no longer fails with an error pointing to atioglxx.dll when exiting the game.
* Enemy Territory: Quake Wars: Corruption is no longer noticed when editing levels.
* Guitar Hero 3: Poor performance is no longer noticed when playing the game on a system containing an ATI Radeon™ HD 2900 series of product and running the Windows Vista operating system.
* Halo 2: Enabling AA in the Catalyst™ Control Center no longer results in the Halo 2 logo flickering when loading or quitting a campaign.
* Halo2: Starting a new campaign with the default Catalyst™ Control Center setting and the default game settings no longer results in texture corruption being noticed when playing the game under Windows Vista with an ATI Radeon™ HD 2100 series of product installed.
* Oblivion: Flickering is no longer noticed when first making a move if the in-game settings are set to maximum, CrossFire™ is enabled and AF is set to 16x.
* Resident Evil 4: Minor flickering is no longer noticed when playing the game on a system running Windows Vista along with CrossFire™ being enabled.
* Splinter Cell Double Agent: Multicolored rectangles are no longer noticed when UBM is enabled on systems containing an ATI Radeon™ HD 34x0 series of product.
* World of Conflict: Text corruption is no longer noticed when playing the game under Windows Vista with CrossFire™ enabled.
* World of Warcraft: Flickering and missing objects are no longer noticed when playing the game in OpenGL mode.
* Enemy Territory: Quake Wars: Corruption is no longer noticed when editing levels.
* World of Warcraft: Flickering and missing objects are no longer noticed when playing the game in OpenGL mode.
* Doom3: Running an OpenGL application with CrossFire™ enabled and AA set to 16x no longer results in corruption being noticed when the display resolution is set to 2048x1536 or higher.
* TianXia2: Launching the game no longer results in the Windows XP operating system rebooting.
* Native CrossFire™ can now be enabled when the primary display device is connected to the master card using a CV or TV.

I think it will be interesting to do the same article with both multi-GPU solutions... The GeForce 9800 GX2 and the Radeon HD 3870 X2... All said and done it looks like they had good intentions, but for me they need newer hardware... Also 8.1 drivers on the Radeon 3870 were the beta drivers, so I don't think it's fair to compare beta drivers on a brand new card to that of the GeForce 8800 Ultra.. A fair test would be to go back to April 28, 2007 and use the Forceware 158.18 drivers that the Ultra came out with... NVIDIA has come a long way since then.

[DMB2000uk]

Yeah, it smelt fishy from the start. The stats pulled to demonstrate the Ultra's 'superiority' looked like choice stats too.

Dan

[Kougar]

Heh, if Crossfire is "broken", then what does that make Nvidia's SLI? People in glass houses shouldn't go throwing GX2's around...

As I recall, ATI was actually besting NVIDIA in regards to multi-GPU scaling. Especially when looking at 3-way setups. A few sites have done them, but Anandtech had several articles focusing on scaling...

[w00fd06]

As I recall, ATI was actually besting NVIDIA in regards to multi-GPU scaling. Especially when looking at 3-way setups. A few sites have done them, but Anandtech had several articles focusing on scaling...

I was gonna say the same, and now the 9800GX2 is out, and it's not exactly a deal, nor is it much more powerful than the 3870X2 considering the price. maybe ATI will be able to stay ahead for a bit.
I also heard about ATI planning something to completely crush nVidia, who knows what it is, probably no one, but rumors are always nice

[Apoptosis]

I also heard about ATI planning something to completely crush nVidia, who knows what it is, probably no one, but rumors are always nice

R700

[skier]

I was gonna say the same, and now the 9800GX2 is out, and it's not exactly a deal, nor is it much more powerful than the 3870X2 considering the price. maybe ATI will be able to stay ahead for a bit.
I also heard about ATI planning something to completely crush nVidia, who knows what it is, probably no one, but rumors are always nice

well, GX2 isnt out just yet, wiki says tomorro is projected, with the 25th for the GTX, and GX2 is "allegedly" 30% more powerful than 8800ULTRA, which the 3870X2 matches with performance, provided this isnt a lie, the GX2 will yeild plenty more performance, but the pricepoint is the big unknown, and isnt R700 only designed to "compete" with G100 which isnt even really planned yet?

[DMB2000uk]

Lets hope it rocks Nvidia's boat

Fingers crossed this card will be from the fruits of the AMD/ATI merger.

Dan

[w00fd06]

R700

oh, well that wasn't obvious enough , is it set to be out this year?

and GX2 is "allegedly" 30% more powerful than 8800ULTRA, which the 3870X2 matches with performance, provided this isnt a lie, the GX2 will yeild plenty more performance, but the pricepoint is the big unknown, and isnt R700 only designed to "compete" with G100 which isnt even really planned yet?

this is what i've heard:

We printed the numbers a few days ago, here if you missed it, and they are pretty woeful. Lets assume the worst for the 3870X2, call it 13,500 3DMark06, and lets assume the best for the the 9800GX2, with a driver drop and other hacks, say it hits 15K. A little math (15,000/13,500 = 1.111) tells us that it will be about 11% faster than the 3870X2. No problem so far.

The funny part? It is going to retail for $599 US, you read that right, $599. The X2 is at $449, 75% of the price. A bit more math tells us that the price performance (3DMark/$) of the X2 is 30/$ and 25/$ for the GX2. That is one hell of a difference, and it is based on best case assumptions. To put it in different terms, you could get a 3870X2 AND a 3870 for a three way for around the same money.