CrossFireX & SLI scaling today

[Kaotik]

http://www.overclockersclub.com/revi...ssfire_vs_sli/

Overclockersclub made a big review on CF & SLI scaling.
They did a big work for it, no doubt, but there's some issues too that stick out:
- 4850X2 CFX beating 4870X2 CFX in more than one case seems odd
- No single chip results at all
- No scaling results, so had to count scaling myself :P

And then some personal quirks on how the review seems quite damn biased:
- PTBoats - game pretty much no-one ever heard of outside benchmarks, known to favor heavily nVidia products
- 2 GPU PhysX games, in which CPUs are doing the heavy GPU PhysX calculations for Radeons, making their results just tied to how fast or slow the CPU can crunch physics, not related to cards performance at all
- No similar cases which would favor ATI over nVidia (like DX10.1 titles in DX10.1 mode, which nV can't run, just as "fair" as the GPU PhysX running on CPU for Radeons, didn't even notice if they used DX10.1 mode in the couple games that support it for Radeons.

Then to the results - absolute FPS numbers can be found behind the link, but here's the scaling numbers for those cards it could be calculated on - GTX295 > GTX925 QuadSLI, HD4870X2 > HD4870X2 CFX, HD4850 > HD4850X2 CFX

Average Scaling:
HD4850: 47,35%
HD4870: 36,11%
GTX295: 35,67%

Average Scaling without the two GPU PhysX games:
HD4850: 52,32%
HD4870: 39,60%
GTX295: 38,12%


And then the individual game results:

Game        	GTX295	4870X2	4850X2
FarCry 2 1920	3,45 %	40,85 %	61,02 %
FarCry 2 2560	31,03 %	77,50 %	74,36 %
Crysis 1920	18,64 %	23,91 %	48,72 %
Crysis 2560	6,78 %	24,39 %	114,81 %
Warhead 1920	81,82 %	11,54 %	50,00 %
Warhead 2560	57,89 %	55,56 %	27,78 %
PTBoats 1920	13,41 %	16,39 %	32,69 %
PTBoats 2560	40,98 %	43,48 %	59,46 %
Bioshock 1920	39,23 %	33,84 %	69,87 %
Bioshock 2560	71,85 %	58,73 %	66,34 %
WiC 1920	22,22 %	1,69 %	50,00 %
WiC 2560	20,00 %	15,00 %	77,42 %
Cryostasis 1920	26,09 %	7,14 %	7,14 %
Cryostasis 2560	25,00 %	7,14 %	7,69 %
IL2 1920	8,53 %	0,00 %	5,06 %
IL2 2560	1,83 %	16,92 %	-1,52 %
CoD IV 1920	34,66 %	39,41 %	69,33 %
CoD IV 2560	68,75 %	73,57 %	88,29 %
CoD WaW 1920	58,82 %	48,62 %	82,02 %
CoD WaW 2560	65,96 %	97,30 %	92,73 %
WH40k 1920	33,96 %	41,49 %	51,76 %
WH40k 2560	50,00 %	68,57 %	87,10 %
FEAR2 1920	44,17 %	24,62 %	18,35 %
FEAR2 2560	66,67 %	85,71 %	74,29 %
Lost P 1920	33,33 %	16,67 %	55,81 %
Lost P 2560	74,00 %	28,57 %	36,36 %
Mirror's 1920	0,00 %	0,00 %	0,00 %
Mirror's 2560	3,45 %	4,35 %	-4,35 %
Fallout 1920	0,00 %	-6,74 %	2,33 %
Fallout 2560	1,11 %	-4,71 %	14,67 %
Dead Space 1920	41,96 %	35,16 %	32,89 %
Dead Space 2560	79,19 %	52,46 %	44,23 %
Left4Dead 1920	12,67 %	33,61 %	46,15 %
Left4Dead 2560	36,73 %	49,37 %	52,24 %
STALKER 1920	26,53 %	36,36 %	46,67 %
STALKER 2560	46,43 %	48,15 %	55,56 %
3DMark06 1920	11,22 %	12,25 %	29,04 %
3DMark06 2560	18,49 %	29,14 %	43,29 %
Vantage 1920	66,33 %	86,37 %	59,75 %
Vantage 2560	83,55 %	109,81 %	64,47 %

--

[Dave Baumann]

Originally Posted by Kaotik

They did a big work for it, no doubt, but there's some issues too that stick out:
- 4850X2 CFX beating 4870X2 CFX in more than one case seems odd

Absolute performance would be a little odd, but if we are talking scaling then I'd expect that as CPU and bus saturation will inhibit 4870 X2's scaling

[Kaotik]

Originally Posted by Dave Baumann

Absolute performance would be a little odd, but if we are talking scaling then I'd expect that as CPU and bus saturation will inhibit 4870 X2's scaling

To clear that out, in absolute FPS the 2x HD4850X2 are beating 2x HD4870X2's in more than one test
edit:
Cases where the 4850X2's are beating 4870X2's:
(presented 2x HD4870X2 - 2x HD4850X2)
Crysis 1920x1200: 57 - 58 FPS
Crysis 2560x1600: 51 - 58 FPS
Crysis Wars 1920x1200: 29 - 33 FPS
World in Conflict 1920x1200: 60 - 69 FPS
World in Conflict 2560x1600: 46 - 55 FPS
Fallout 3 1920x1200: 83 - 88 FPS
Fallout 3 2560x1600: 81 - 85 FPS
3DMark06 1920x1200: 24879 - 24910 3DMarks

Also, in IL1946 at 2560x1200 single 4850X2 is 1 FPS faster than single 4870X2,

[lopri]

Originally Posted by Dave Baumann

Absolute performance would be a little odd, but if we are talking scaling then I'd expect that as CPU and bus saturation will inhibit 4870 X2's scaling

Do you mean x16/x16 PCIe? That's not enough for 4870X2's?

[Sxotty]

Still Dave's point stands. And it also explains why gtx 295 scaling is so crappy. If they tested 275 scaling, 265 scalling it would probably look better than 295 scaling regardless of whether the FPS actually passed. And they still might be about the same and within margin of error as 1 FPS likely is in one of your examples.

[OpenGL guy]

Originally Posted by Kaotik

To clear that out, in absolute FPS the 2x HD4850X2 are beating 2x HD4870X2's in more than one test
edit:
Cases where the 4850X2's are beating 4870X2's:
(presented 2x HD4870X2 - 2x HD4850X2)
Crysis 1920x1200: 57 - 58 FPS
Crysis 2560x1600: 51 - 58 FPS

4850x2 gets 27 fps yet two 4850x2s score 58? Seems like something is wrong with the testing or results here.

3DMark06 1920x1200: 24879 - 24910 3DMarks

This is not a significant difference, the run-to-run variance for this benchmark is higher than that.

[lopri]

Oh I am not comparing the results from the review. I'm more curious about the necessary bandwidth for these dual-chip cards. I am skeptical that x16/x16 is saturated by either 295's or X2's, as long as the motherboard is properly built. I was thinking even x8/x8 configuration wouldn't hinder CrossFireX or Quad-SLI. At least from a theoretical point of view. I thought the overhead occurred not from GPU-GPU, but rather drivers-CPU-system memory (and back to PCIe/GPU).

[ChrisRay]

PCIE bandwith? No. Its a myth that you need fast CPU to get nearly 90% scaling from SLI. And honestly the bridge chip is handling most of that anyways.

1) You can get great scaling from multi GPU AFR solutions providing that your Pixel Fillrate or Z-Fillrate are your primary bottlenecks. So extreme AA levels of AA, Transparency SS, Highest resolutions. Which is why multi GPU testing is always done at high resolutions.

2)CPU limited scaling occurs in highly geometry limited areas. And I suspect this is partially due to vertex setup in alot of games. In situations like this. You aren't gonna see a major difference between low end multi GPU and high end multi GPU. Low resolutions ect will show this alot.

Just an example. With Quad GTX 295 and say a commonly CPU limited game such as WoW. With 16xAA/16xAF with transparency SS will give zero scaling. Yet a dual 8800GT card will show perhaps 40-50% scaling at the same settings. But if you set both cards to say 32xS AA with transparency AA. A GTX 295 will scale from 100% ((single GPU) to 375%-380% ((4 GPU)) and the 8800GT SLI setup will scale roughly 100%(single GPU to 190% (Dual GPU). While the scaling comparisons shown here ((percentages)) are interesting foot notes. They dont neccasarily represent how software/hardware will scale at the right settings.

[XMAN26]

Kaotik, you wanna fault them for the none use of DX10.1, fine do so as I can concur to this, it didn't make sense not to. But dont fault them for using PhysX enabled games, blame ATI/AMD for not getting on the ball to get something to the public to counter it yet. Also keep in mind, if they make Havoc run thru OCL, Nv is supporting this aswell and dont be surprised if they dont do Havoc thru Cuda.

[Kaotik]

Originally Posted by XMAN26

Kaotik, you wanna fault them for the none use of DX10.1, fine do so as I can concur to this, it didn't make sense not to. But dont fault them for using PhysX enabled games, blame ATI/AMD for not getting on the ball to get something to the public to counter it yet. Also keep in mind, if they make Havoc run thru OCL, Nv is supporting this aswell and dont be surprised if they dont do Havoc thru Cuda.

Tell me even one reason to use GPU PhysX settings in benchmark comparing performance of graphics cards platforms when you know that one of them is going to perform only as fast as your CPU can crunch the physics meant for nV GPUs
It would be the same as running benchmarks, like I used as example, in DX10.1 mode only and say "whoopsie, nV cards can't run this at all, let's use it anyway and use reference rasterizer numbers for nV cards"
It doesn't make any sense.

[XMAN26]

Originally Posted by Kaotik

Tell me even one reason to use GPU PhysX settings in benchmark comparing performance of graphics cards platforms when you know that one of them is going to perform only as fast as your CPU can crunch the physics meant for nV GPUs
It would be the same as running benchmarks, like I used as example, in DX10.1 mode only and say "whoopsie, nV cards can't run this at all, let's use it anyway and use reference rasterizer numbers for nV cards"
It doesn't make any sense.

If the games in question, I dont own any, have the option to turn off physx, then they should do that and make the notation to the bench that ATI is not running physx of any kind and Nvidia is. Same thing could be done for games that do DX10.1.

But again, most of teh blame still falls on AMD/ATI for not getting anything out in any kind of support for havok as of yet on the shelves in the form of a game. Nvidia on the other hand spends tons to help developers program and use physx.

[willardjuice]

Nvidia on the other hand spends tons to help developers program and use physx.

Oh I would love to hear the long list of developers that are lined up to use (gpu) PhysX...

[XMAN26]

Originally Posted by willardjuice

Oh I would love to hear the long list of developers that are lined up to use (gpu) PhysX...

Given the number of games out now that use it and coming out, all by different developers, do you really need a list?

http://www.nzone.com/object/nzone_physxgames_home.html
Game Title Developer Platform
2 Days to Vegas Steel Monkeys PC
Adrenalin 2: Rush Hour Gaijin Entertainment PC
Age of Empires III Distineer Studios PC, Mac
Age of Empires III: The WarChiefs Distineer Studios Mac
Alpha Prime Black Element Software PC
Auto Assault Net Devil PC
Backbreaker Natural Motion TBA
B.A.S.E. Jumping Digital Dimension Development PC
Bet on Soldier: Blackout Saigon Kylotonn Entertainment PC
Bet on Soldier: Blood of Sahara Kylotonn Entertainment PC
Bet on Soldier: Blood Sport Kylotonn Entertainment PC
Beowulf Ubisoft PS3, X360
Bladestorm: The Hundred Years' War Koei PS3, X360
Captain Blood Akella PC, X360
Cellfactor: Combat Training Artifical Studios, Immersion Games PC
Cellfactor: Revolution Artifical Studios, Immersion Games PC
City of Villains Cryptic Studios PC
Crazy Machines II FAKT Software PC
Cryostasis Action Forms PC
Dark Physics The Game Creators PC
Desert Diner Tarsier Studios PC
Dragonshard Atari PC
Dusk 12 Orion PC
Empire Above All IceHill PC
Empire Earth III Mad Dog Software PC
Entropia Universe MindArk PC
Fallen Earth Icarus Studios PC
Fury Auran Games PC
Gears Of War Epic Games PC, X360
Gluk'Oza: Action GFI Russia PC
GooBall Ambrosia software Mac
Gothic 3 Piranha Bytes PC
Gunship Apocalypse FAKT Software PC
Heavy Rain Quantic Dream PC
Hero's Jorney Simutronics PC
Hour of Victory nFusion Interactive X360
Hunt, The Orion PC
Huxley Webzen, Inc PC, X360
Infernal Metropolis Software PC
Inhabited island: Prisoner of Power Orion PC
Joint Task Force Most Wanted Entertainment PC
Kuma\WAR Kuma Reality Games PC
Magic ball 3 Alawar Entertaiment PC
Mass Effect BioWare PC, X360
Medal of Honor: Airborne EA Los Angeles PC, X360
Metro 2033 4A Games PC
Mobile Suit Gundam: Crossfire BEC PS3
Monster Madness: Battle for Suburbia Artificial Studios PC, X360
Monster Truck Maniax Legendo Entertainment PC
Myst Online: URU Live Cyan Worlds PC
Open Fire BlueTorch Studios PC
Paragraph 78 Gaijin Entertainment PC
Pirates of the Burning Sea Flying Lab Software PC
PT Boats: Knights of the Sea Akella PC
Rail Simulator Kuju Entertainment Ltd PC
Red Steel Ubisoft Paris Wii
Rise Of Nations: Rise Of Legends Big Huge Games PC
Roboblitz Naked Sky Entertainment PC, X360
Sacred 2 ASCARON Entertainment PC
Sherlock Holmes: The Awakened Frogwares Game Development Studio PC
Showdown: Scorpion B-COOL Interactive PC
Silverfall Monte Cristo PC
Sovereign Symphony Ceidot Game Studios PC
Sonic and the Secret Rings SEGA Wii
Speedball 2 Kylotonn Entertainment PC
Stalin Subway, The Orion PC
Stoked Rider: Alaska Alien Bongfish Interactive Entertainment PC
Switchball Atomic Elbow PC
Tension Ice-pick Lodge PC
Tom Clancy's Ghost Recon Advanced Warfighter GRIN PC, X360
Tom Clancy's Ghost Recon Advanced Warfighter 2 GRIN, Ubisoft Paris PC, X360
Tom Clancy's Rainbow Six Vegas Ubisoft Montreal PC, PS3, X360
Tom Clancy's Splinter Cell: Double Agent (multiplayer) Ubisoft Shanghai PC, X360
Tortuga: Two Treasures Ascaron Entertainment PC
Two Worlds Reality Pump PC
Ultra Tubes Eipix PC
Unreal Tournament 3 Epic Games PC, PS3, X360
Unreal Tournament 3: Extreme Physics Mod Epic Games PC
Warfare GFI Russia PC
Warmonger: Operation Downtown Destruction Net Devil PC
W.E.L.L. Online Sibilant Interactive PC
Winterheart's Guild Zelian Games PC, X360
WorldShift Black Sea Studios PC

What would be nice is for a list of titles/devs working on Havoc in games.




































































Waits for the crickets to stop chirping.

[willardjuice]

Phew for a second I thought you were actually going to answer my question.

[XMAN26]

Originally Posted by willardjuice

Phew for a second I thought you were actually going to answer my question.

If you bothered with the link, most devs have signed on for Physx.

Biggies in the list
Atari
EA
Bioware
Epic
Orion
Ubisoft
Sega

And I wouldn't be surprised in the least when ID joins up.

[willardjuice]

Originally Posted by XMAN26

If you bothered with the link, most devs have signed on for Physx.

Biggies in the list
Atari
EA
Bioware
Epic
Orion
Ubisoft
Sega

I'm talking about games (and real games, not like Cellfactor) that actually have some of their physics calculations accelerated by the gpu. And no situations like Epic/UT3 do not count since only certain 3rd party maps (made by Ageia) are accelerated. When you actually analyze the circumstances you'll find hardly any developer support for gpu physics at the moment.

Originally Posted by XMAN26

And I wouldn't be surprised in the least when ID joins up.

Yeah that's not going to happen.

[XMAN26]

Originally Posted by willardjuice

I'm talking about games (and real games, not like Cellfactor) that actually have some of their physics calculations accelerated by the gpu. And no situations like Epic/UT3 do not count since only certain 3rd party maps (made by Ageia) are accelerated. When you actually analyze the circumstances you'll find hardly any developer support for gpu physics at the moment.




Yeah that's not going to happen.

You still have not bothered with the link or even the lsit of games. There areseveral games in the list by the developers. Epic has more than 2(other than UT3). And keep in mind any Ageia affects support writen into games should count, they also get run on the GPU now.

[willardjuice]

Originally Posted by XMAN26

You still have not bothered with the link or even the lsit of games. There areseveral games in the list by the developers. Epic has more than 2(other than UT3).

What Epic game supports PhysX via the gpu?

(spoiler alert: none of them do)

[Betanumerical]

Originally Posted by willardjuice

What Epic game supports PhysX via the gpu?

(spoiler alert: none of them do)

Whilst not a EPIC game I think its worth a mention that Mirrors Edge supports it and uses UE3.

[XMAN26]

Originally Posted by willardjuice

What Epic game supports PhysX via the gpu?

(spoiler alert: none of them do)

Really? You can't run Physx on UT3 on the GPU? There isn't DLable content with the same? Regardless of your own view, it is still physx on a GPU.
Then there is GRAW2 which has added Physx support, AN ACTUAL GAME BY YOUR OWN STANDARDS!

[willardjuice]

There isn't DLable content with the same?

There's three 3rd party maps that no one plays, congratulations.

Then there is GRAW2 which has added Physx support, AN ACTUAL GAME BY YOUR OWN STANDARDS!

Yeah no where did I say PhysX games (via gpu) do not exist. As you and Beta so astutely pointed out, they do in fact exist. My point is that market (at the moment) is extremely minimal.

[ChrisRay]

Originally Posted by XMAN26

Really? You can't run Physx on UT3 on the GPU? There isn't DLable content with the same? Regardless of your own view, it is still physx on a GPU.
Then there is GRAW2 which has added Physx support, AN ACTUAL GAME BY YOUR OWN STANDARDS!

UT3 Physics ((with UT3 Physics Pack)) does run off the GPU. But that Physics Pack is the only part of the game that makes substancial use of Physics. And will work on PPU and GPU. The rest of the game does not. It is however a really cool PhysX demo. I love the heatray map with all the hail.

Also Sacred 2 in that list does GPU Physics as does Cryostasis.


http://www.techreport.com/articles.x/16820/7

[Kaotik]

Cryostasis and Mirror's Edge are the only 2 titles which actually make any real difference on if you have GPU PhysX or not, Sacred 2.. well, okay, I supposed it could be counted aswell, though some reviewers have said that if you don't know what GPU PhysX is supposed to do in it, you'd never notice it
And XMAN, yes, you can turn PhysX "off" or to "light cpu physics" in all of them

As for Havok titles, here's some for you (I limited the list to games available for PC too, also, the list only includes titles released after 2006 and is NOT including upcoming titles using it like Alan Wake):
Alone in the Dark
Ant Bully
Auto Assault
BioShock
Blacksite : Area 51
Bottle Buster
Company of Heroes
Company of Heros: Opposing Fronts
Dark Messiah of Might and Magic: Elements
Desperados 2: Cooper's Revenge
Destroy All Humans! 2
Dungeons & Dragons Online: Stormreach
F.E.A.R. Files
Fallout3
Full Spectrum Warrior
Full Spectrum Warrior: Ten Hammers
Guitar Hero III
Half-Life 2: The Orange Box
Halo 2
Happy Feet
Harry Potter and the Order of The Phoenix
Hellgate: London
IHRA Drag Racing: Sportsman Edition
Iron Man
Just Cause
Kane & Lynch: Dead Men
LOTR Online: Shadows of Angmar
Mercenaries 2
Micro Machines V4
Over the Hedge
Painkiller
Painkiller: Overdose
Requiem: Bloodymare
Shadowrun
Shrek SuperSlam
Shrek the Third
Soldier of Fortune: Payback
Stranglehold
Superman Returns: The Videogame
The Elder Scrolls IV: Oblivion
The Godfather
The Golden Compass
The Incredible Hulk
Timeshift
Tom Clancy's Splinter Cell Double Agent
Tony Hawk's Proving Ground
World in Conflict
X-Men: The Official Game

edit:
As side note, it seems Havok has been used even in several movies high profile movies: http://www.havok.com/content/view/644/111/

[Kaotik]

One other suspicious issue in the review - check the power consumption page - there's no way HD4870X2 would consume over 100W more than 2x HD4870 1GB's in CF

[XMAN26]

Originally Posted by Kaotik

Cryostasis and Mirror's Edge are the only 2 titles which actually make any real difference on if you have GPU PhysX or not, Sacred 2.. well, okay, I supposed it could be counted aswell, though some reviewers have said that if you don't know what GPU PhysX is supposed to do in it, you'd never notice it
And XMAN, yes, you can turn PhysX "off" or to "light cpu physics" in all of them

As for Havok titles, here's some for you (I limited the list to games available for PC too, also, the list only includes titles released after 2006 and is NOT including upcoming titles using it like Alan Wake):
Alone in the Dark
Ant Bully
Auto Assault
BioShock
Blacksite : Area 51
Bottle Buster
Company of Heroes
Company of Heros: Opposing Fronts
Dark Messiah of Might and Magic: Elements
Desperados 2: Cooper's Revenge
Destroy All Humans! 2
Dungeons & Dragons Online: Stormreach
F.E.A.R. Files
Fallout3
Full Spectrum Warrior
Full Spectrum Warrior: Ten Hammers
Guitar Hero III
Half-Life 2: The Orange Box
Halo 2
Happy Feet
Harry Potter and the Order of The Phoenix
Hellgate: London
IHRA Drag Racing: Sportsman Edition
Iron Man
Just Cause
Kane & Lynch: Dead Men
LOTR Online: Shadows of Angmar
Mercenaries 2
Micro Machines V4
Over the Hedge
Painkiller
Painkiller: Overdose
Requiem: Bloodymare
Shadowrun
Shrek SuperSlam
Shrek the Third
Soldier of Fortune: Payback
Stranglehold
Superman Returns: The Videogame
The Elder Scrolls IV: Oblivion
The Godfather
The Golden Compass
The Incredible Hulk
Timeshift
Tom Clancy's Splinter Cell Double Agent
Tony Hawk's Proving Ground
World in Conflict
X-Men: The Official Game

edit:
As side note, it seems Havok has been used even in several movies high profile movies: http://www.havok.com/content/view/644/111/

Nice to see some games on there, now how about those that actually have Havok ding something for the game itself much like how Physx is doing it for Cryostasis, Mirros Edge, GRAW2 for example.