Rage3D NVIDIA GeForce 7950 GX2 Review
By Mark "Ratchet" Thorne - ratchet@rage3d.com
June 20th, 2006
[ Print ] - [ Close ]
Introduction

NVIDIA GeForce 7950 GX2
NVIDIA GeForce 7950 GX2
So you got yourself a nice new multi-GPU setup and now you think you’re the new King of the Gaming World. You think your machine is so fast you put on a neck-brace and crash helmet before firing up Oblivion. Your desk has speed brakes and a parachute, your chair an ejection system in case things get too hairy too fast. So un-Godly fast is your machine you believe the Pope should drop by and exorcize it of its demonic ways. Birds sing and dolphins leap, all in praise of your 3DMark score. Nothing, not of this universe or the next, could possibly put as many pixels on screen like the machine that is warming your toes right now.

NVIDIA GeForce 7950 GX2
NVIDIA GeForce 7950 GX2
Not so, my FPS addicted friend, not so.

You see, while you were off prancing around the forests of Tamriel, NVIDIA released a new graphics card for those like you, those addicted to the eye-candy and the almighty FPS counter. Not just a graphics card mind you, but really two graphics cards in one. Two PCBs bolted together, two fans, two GPUs, two banks of memory; all in one sneaky little SKU NVIDIA likes to call the GeForce 7950 GX2.


The idea of combining multiple GPUs on one SKU isn’t new; dreams of multi-GPU graphics cards have been floating around almost as long as 3D graphics hardware has.

Advertisement
ATI took a shot at a dual-chip board when they produced and shipped the Rage Fury MAXX, built using two of their Rage Fury 128 Pro chips. Unfortunately, while the MAXX worked and did what it was designed to do (at least under Windows 98), it met up against NVIDIA’s GeForce juggernaut and, even with two chips, simply could not compete. Adding to the demise of the MAXX name was also the fact that it had near fatal problems under Microsoft’s new Windows 2000 OS that ATI never could solve.

Probably the most famous attempt at the idea however was by 3dfx when they started design on the Voodoo 5 line. The Voodoo 5 5000, Voodoo 5 5500, and Voodoo 5 6000 each featured multiple chips on one board using 3dfx’s VSA-100 architecture. The 5000 and 5500 both featured dual-chip designs and the Voodoo 6000 was designed, maybe over-designed, with an amazing four chips on one board. However, the 5000 and 6000 were both eventually cancelled and the Voodoo 5500, while it did ship, met with overwhelming competition from NVIDIA's GeForce 2 and ATI’s new Radeon. This spelled the end for the Voodoo 5 and, ultimately, 3dfx itself.

We’ve also recently saw several partner companies try it. Guys like Gigabyte, ASUS, HIS, and Powercolor have shown and produced some products, with varying degrees of success, using technology from both sides of the Graphics War frontline.

With NVIDIA’s 7950 GX2 however, we have the first instance of a single consumer level card taking the multi-PCB approach to the idea of multi-GPU processing and, with development at the IHV level and proven technology behind it, this may be the first time that a multi-GPU card has real potential for success.



The 7950 GX2

Sporting 48 pixel pipes, 16 vertex shaders, memory bandwidth exceeding 76 GB/s, a pixel-fillrate of 16,000 MP/s and a texel fillrate of 24,000 MT/s; the 7950 GX2, as a “single” card, simply can’t be touched today. Our benchmarks later on will prove that.

The real kick in the pants, however, is that the 7950 GX2 is fully capable of running in SLI with another 7950 GX2. Putting two of them together in a single system gives you what NVIDIA calls Quad SLI, and with four GPUs running at the same time I can’t think of a more appropriate name for it. Some might not agree to look at Quad SLI specs the way NVIDIA does, or even the 7950 GX2 specs for that matter, but just for fun think about a graphics sub-system that has 96 pixel pipes, 32 vertex shaders, and provides 154 GB/s of memory bandwidth over a total of 2GB of graphics memory. If specs made people cool, Quad SLI would turn Steve Urkel into Steve McQueen.

Advertisement
The only problem is that Quad SLI is currently only available at the System Integrator level. If you want it, you have to buy an entire Quad SLI equipped system.  It’s coming to the end-user, but as of yet NVIDIA has a flimsy ETA of “this year” for the necessary drivers to make it all work.

The 7950 GX2 seems pretty spectacular on paper, and Quad SLI would most definitely smoke anything out there, but there is a catch.

You see, at the heart of the GX2 is NVIDIA’s SLI technology. To put it simply, the GX2 is basically two 512MB GeForce 7900 cards stacked one on top of the other and linked together via a sort of expanded SLI bridge. There is no revolutionary GPU hiding under the twin coolers, it’s essentially the same G71 with all the same features that we already have on the 7900 GTX and 7900 GT. And because its heart is SLI, it has all the same disadvantages that come with a regular dual-card SLI setup.

For example, to enable dual-display mode, you have to disable the multi-GPU mode in NVIDIA’s control panel. Not a big deal in all honesty, but it can be a pain if you have dual displays. Also, there are games out there that have problems with SLI and simply don’t work. This can be especially true of older games or even new games that haven’t gotten the attention required from NVIDIA’s developers. In those cases, the 7950 GX2 can’t make use of the second GPU and would end up performing roughly on par with a single 7900 GT.

NVIDIA has gotten SLI to work pretty well over the last couple years since its introduction and it’s pretty painless these days, but those disadvantages still exist and will undoubtedly affect some users from time to time. For the most part, however, those disadvantages are pretty minor when compared to what SLI and the 7950 GX2 offers overall.
A Closer Look at the GX2

As you’ve hopefully noticed by now, the 7950 GX2 consists of two PCBs attached together with a GPU and 512MB of GDDR3 on each half. The top half of the GX2 holds the DVI ports and the auxiliary 6-pin power, while the bottom half connects to the motherboards PEG slot.



Each GPU is cooled by a single-slot cooler, but obviously the GX2 as a whole is a dual-slot card. Even with two fans running, the coolers themselves are very quiet, even at the higher settings, and do a decent enough job of keeping the GPUs cool. Unlike the 7900 GX2 cooling design, there is no cutout in the top PCB to help air get to the bottom coolers fan.

Connecting the two halves of the GX2 is a modified SLI bridge. It does the same job of data transfer and synchronization as a traditional “over-the-top” SLI bridge, but also has an added job of balancing the power distribution between the two halves of the GX2.



Also noteworthy is the PCI-Express switch hidden behind the cooler on the bottom card. It provides a 16 lane connection to the system, and two 8 lane connections for the GPUs to talk in a peer-to-peer fashion over the SLI link. With the older 7900 GX2 GPU-to-GPU communications was routed over 16 lanes (for 48 total lanes), but to reduce the overall length of the 7950 GX2 NVIDIA changed it to 8 lanes.

PCI-Express Switch
PCI-Express Switch
The main reason for having this switch is to permit wider motherboard compatibility. Intel based boards and even ATI boards (like the A8R32-MVP this reviewer is using) are on NVIDIA’s list of compatible motherboards, and more boards are added as they receive the proper BIOS update from the manufacturers.

Just to the right of the PCIE switch is the G71 GPU, one half of the power behind the GX2. NVIDIA calls this A2 stepping of the G71 simply the “7950” GPU in all their documentation and correspondence. The new A2 stepping and name certainly does indicate possible improvements over the G71 GPU found on the 7900 GT and GTX, and I’ve got it on pretty good authority that at least some of those improvements are power related.

The memory, 512MB of GDDR3 per card, also hidden beneath the coolers, is Samsung model K4J523240Q-BC14 rated for 1.4ns or 700MHz.

Top PCB
Top PCB
Bottom PCB
Bottom PCB
G71 GPU
G71 GPU
1.4ns Samsung GDDR3
1.4ns Samsung GDDR3

The GX2 also features two Dual-Link DVI ports, common these days on high-end video hardware, but which also happily support HDCP. It seems silly to make a happy note about having support for technology that is designed to restrict the use of something bought and paid for, but without HDCP support on our graphics cards we couldn’t watch future HD-DVD and Blu-ray content in all its intended HD glory on our systems—score one for the MPAA there I guess.



Test Setup

Resolutions

Image Quality Settings

Test System Specs

  NVIDIA 7950 GX2 BFG 7900 GTX OC ATI X1900 XTX
Core G71 x 2 G71 R580
Silicon Process 90nm 90nm 90nm
Transistor Count
(millions)		 556
(278 x 2)
 278 384
Core Speed MHz 500 670 650
Memory Speed MHz (Effective) 600
(1.2 GHz)
 820
(1.640 GHz)
 775
(1.55 GHz)
Memory Size 1,024 MB
(512 MB x 2)
 512 MB 512 MB
Bus Standard PEG 16x PEG 16x PEG x16
Bus Width 512bit
(256bit x 2)
 256bit 256bit
ROPs 32
(16 x 2)
 16 16
Pixel Shaders 48
(24 x 2)
 24 48
Vertex Shaders 16
(8 x 2)
 8 8
Peak Memory Bandwidth
(GB/s)		 76.8
(38.4 x 2)
 51.2 49.6
Pixel Fillrate
(million pixels/sec)		 16,000
(8,000 x 2)
 10,400 10,400
Texel Fillrate
(million texels/sec)		 24,000
(12,000 x 2)
 15,600 10,400
API Compliancy DX 9.0c
Shader Model 3.0		 DX 9.0c
Shader Model 3.0		 DX 9.0c
Shader Model 3.0

Games Benchmarks (click for settings)

The Windows XP desktop was set to 1280x960 with a 32bit color depth and 85Hz refresh rate for all tests. Refresh rate locks for 3D graphics modes, as supported by both NVIDIA and ATI graphics control panels, was not enabled. V-Sync was forced off via the graphics card control panel as well. All other graphics card control panel settings were left to their default settings unless otherwise noted.

Anti-Aliasing and Anisotropy were applied in the game engine where the options existed. For games that did not support those options natively, the graphics card control panel was used.

Custom batch files were used when possible for automated benchmarking (the details of the commands used are outlined for each test). When manual benchmarking was necessary Fraps was used.

Benchmarking was done with Windows set to the "Adjust for best performance" profile, and all unnecessary Windows services and hardware devices were disabled. The latest drivers for each necessary hardware component were installed prior to testing and kept consistent throughout.

Sound and networking interfaces were enabled for all tests.

To setup the test machine we installed Windows XP, patched and tweaked it, and installed all the required games, apps, utilities, and hardware drivers needed for the testing procedure except for the graphics drivers. Using Norton Ghost 2003, we then cloned the drive onto a second identical hard-drive. After that we installed the ATI drivers on one hard-drive and the NVIDIA drivers on the other. Testing the videocards was then a simple matter of swapping videocards and hard-drives when required.

Benchmarks

Half-Life 2: Episode 1

To get results from Episode 1 we recorded two custom timedemos from the citadel_03 and the c17_02a levels and averaged the results to get the final score. citadel_03 is from the core reactor level early in the game and contains a lot of HDR effects while c17_02a contains game action with a lot of explosions and enemies. The benchmark procedure is similar to the one for Half-Life 2 on the next page.



Ghost Recon: Advanced Warfighter

To benchmark Ghost Recon we loaded a saved game and used Fraps to log framerates during a gameplay sequence for about 45 seconds. To get accurate results we made five recordings per resolution and then averaged the results to get the final score.



The Elder Scrolls IV: Oblivion

To benchmark Oblivion we loaded two saved game and used Fraps to log framerates during gameplay sequences for about 45 seconds. One of the saved games was in a wilderness environment with a lot of foilage and the other was in a mountainous environment with little foilage. To get accurate results we made five recordings per resolution, per saved game, and then averaged the results to get the final score.



F.E.A.R.

Benchmarking FEAR was simply a matter of running the in-game Performance Test. Some of the action in the sequence is random, but for the most part it produces reliable, repeatable results. It doesn't reflect actual game play but it should give us an idea of how these cards will perform relative to each other.



Battlefield 2

Battlefield 2 benchmarking is a little tricky. It has a built in time demo feature, but the results it produces can be very unreliable because it starts logging frame rate on the menu screen, before the actual demo starts. The most reliable method to get results is to take the frame rate log the time demo produces (.csv file) and sample the last few thousand frames (in our case we sampled the last 7000 frames). You need to sample the last of the frames because the game starts logging timedemo results duing the loading sequence, which greatly skews the results.. I used a custom timedemo to get the results below.



Half-Life 2: Lost Coast

To get results from Episode 1 we recorded two custom timedemos from the beginning of the map and near the end and averaged the results to get the final score. The benchmark procedure is similar to the one used for Half-Life 2 on the next page.

Benchmarks (II)

GT Legends

To get GT Lengends results we played back a replay and used Fraps to log the framerate for 90 seonds. Anti-Aliasing was set via the GT Legends configuration utility where we used "Level 3", Anisotropic was set via the graphics card control panel.



Splinter Cell: Chaos Theory

Splinter Cell Chaos Theory was also benchmarked using a custom batch file. Anti-aliasing and anisotropy were set within the game. The first chart shows all three cards using the same Shader Model 1.1 path, which is the setting the game defaults to.

In the second chart the Shader Model 3.0 path was chosen and all the SM3.0 features were enabled. Even though the X1000 family supports HDR with AA, Splinter Cell Chaos Theory needs to be patched for it to work. As yet there is no such patch, so we only have 0/0 and 0/16 results there.



Half-Life 2

Testing of Half-Life 2 was done using 4 custom Source Engine 7 time demos from various sections of the game. Because there is some frame rate variance during the Half-Life 2 benchmark process, we ran each timedemo for each resolution and AA/AF setting 3 times then averaged the results to get the final score. Anti-aliasing and Anisotropy were set on the command line.

A batch file was used to automate testing; the command line is below for reference. This batch file was used for each card that was tested. The settings surrounded by < > change for each pass:

"hl2.exe" +r_fastzreject 1 +r_waterforcereflectentities 1 -novid -nosound -width <resolution width> -height <resolution height> +mat_antialias <anti-aliasing> +mat_forceaniso <anisotropy> +mat_trilinear 1 +timedemoquit <timedemo>

hover your mouse over the yellow text to get a brief description of what each switch does

 



Doom 3

Doom3 was tested with 3 custom timedemo and the resutls of each timedemo were averaged to get the final score for that resolution and setting. We benchmarked combinations of Anti-Aliasing and Anisotropy over the resolutions shown in the chart below. Anti-aliasing and Anisotropy were set on the command line.

Another batch file was used to automate Doom 3 testing as well. The command line is below for reference. This batch file was used for each card that was tested. The settings surrounded by < > change for each pass:

"doom3.exe" +set logFile 1 +set com_showFPS 1 +set r_multiSamples <anti-aliasing> +set r_mode <resolution mode> +set image_anisotropy <anisotropy> +set timescale 7 +playdemo <demoname> +wait 1000 +timedemoquit <demoname>
hover your mouse over the yellow text to get a brief description of what each switch does



Far Cry

Again a custom batch file was used to benchmark Far Cry. We used 3 demos included with the newer patches to test performance (from the Cooler, Training, and Volcano levels). Because Far Cry benchmark frame rates can vary between each subsequent pass (sometimes fairly significantly), I ran each of the 3 demos 3 times, then averaged the 9 results to get the final score for the detail level.

The command line for the batch file I used to automate Far Cry benchmarking is below. This batch file was used for each card that was tested. The settings surrounded by < & > change for each pass:

"farcry.exe" -DEVMODE "s_soundEnable 0" "r_width <horizontal res>" "r_height <vertical res>" "demo_num_runs 2" "map <mapname>" "demo <mapname>" "demo_quit 1" "r_FSAA <1|0>" "r_FSAA_samples <AA Samples>" "r_Texture_Anisotropic_Level <AF Level>"
hover your mouse over the yellow text to get a brief description of what each switch does

Overclocking

To overclock the GX2 we enabled Coolbits via the registry.

I let Coolbits find the optimal overclock, then manually tweaked it as high as Coolbits would let it go. The core overclock was a respectable 100MHz over the stock settings, but amazingly the memory overclocked 200MHz, maxxing out the Coolbits slider. It's likely the memory could have gone even higher.

Core Speed 606MHz, Memory Speed 800MHz (1.6GHz)
Core Speed 606MHz, Memory Speed 800MHz (1.6GHz)

Performance results using Splinter Cell Chaos Theory are shown below:

[ No AA / No AF ] [ No AA / 16x AF ]
highend1
Conclusion

I would like to say I was extremely impressed with the 7950 GX2 and all that it offers, and in most cases I am. It’s undeniably fast. It can tear through games at an unprecedented clip, even making a bottleneck out a FX-60 CPU in many cases. High resolutions and settings are what this card loves, and it makes things like Anti-Aliasing something that you can leave on and forget about.

It’s also quiet, which is surprising for a card with two fans. Even under load with the fans spinning up the thing is about as quiet as a church mouse.

Overclocking was also extremely impressive. Anytime you hit the limits on an overclocking tool, without even really trying, you know you’re onto something good (makes you wonder why the thing wasn’t clocked even higher than it is, really).

Unfortunately, NVIDIA still needs to do a little work on the drivers. I did experience some crashing during various games at random intervals, even after I changed from the 91.29 betas that were provided to reviews to the newer public 91.31 betas.

I also experienced a few performance anomalies which are noticeable in F.E.A.R. with AA enabled and, to a lesser extent, with Battlefield 2 benchmarking. There were some other cases where I had to struggle with the drivers--uninstalling, cleaning, and reinstalling--to get accurate results as well. In contrast, the 7900 GTX and X1900 XTX testing procedures went smooth.

The GX2 is unquestionably the fastest card currently available. There’s no other single card out there that can touch it. It makes me wonder though: why does this thing even exist? NVIDIA’s own 7900 GTX was and still is an extremely impressive part and easily a match for ATI’s X1900 XTX flagship, so it’s not like NVIDIA was desperately playing catch up. The GX2 doesn’t introduce DX10 capabilities, or HDR+AA, or even angle-independent filtering. In fact, it offers no new features over the 7900 GTX at all. Was the GX2 a design challenge, an opportunity maybe for NVIDIA engineers to get their feet wet with a multi-GPU SKU? Or, more likely, was it designed simply to lay claim to the Speed crown, bragging rights that marketing departments hold so dear and that allow them to sell so many cards? It certainly seems that way, and if speed and the great FPS counter is what you are primarily interested in, the GX2 definitely delivers. If next gen features and image quality are what you’re looking for then there’s still some waiting to do as the 7950 GX2 doesn’t deliver anything in that department.

 

Verdict

Ultimately, what we have here is the world’s fastest single graphics card. There is no debate, no argument, and no other choice if what you want is framerate.

Pros

Cons