Tech-Hounds.com

Revisiting SLI

Admittedly, we are skeptical of new technologies and products. After all, we think a healthy dose of skepticism is good since that means we reserve judgment until we saw proof to the contrary. When we first looked at what multi rendering has to offer with SLI, we were dismayed to say the least. From our experience, SLI doesn't always make good on its promise of performance. But at least, SLI have made good on the promise of quality, every time with every game.

Seeing how successful NVIDIA is with SLI, the once critic of consumer multi rendering solution, rival ATI, decided they want a piece of the pie as well. Though reviewers generally felt ATI's Crossfire was late to do any good for their aging X800 and X850 cards, most agree the platform has the potential to rival SLI. With their latest X1K series of cards, Crossfire seems to deliver on both performance and quality. The only thing they need now is a much simpler setup, which already made its way to the mainstream parts.

Faced with this new threat from the Canadian manufacturer, NVIDIA answered with fervor in a two headed assault - Quad SLI and the GeForce 7950 GX2. Undoubtedly, these are meant for the high end market. However, the more interesting update from NVIDIA is actually under the hood of the updated GeForce 7600 and 7900 series. The GeForce 7600 family now comes with a much higher clocked core, 5 vertex shader units, 12 pixel shader units and 8 ROP units. In comparison, the GeForce 6600 series comes with 3 vertex shader units, 8 pixel shader units and 4 ROP units. On the high end of the spectrum, the GeForce 7900GTX is now running at a much higher clock, very close to ATI's Radeon X1900XTX. Most of the design of the GeForce 7800GTX is left intact (8 vertex shader units, 24 pixel shader units, 24 ROP units). NVIDIA is determined to win the mainstream parts battle.

We have reviewed these high end parts, but not the mainstream parts. It's not until this month, that we will be publishing our comparison of various GeForce 7600 cards. Since we now have several GeForce 7600GT cards in our lab, we think it's time to see what SLI has to offer with these new cards. Most of our experience with SLI have been with older cards, the GeForce 6600 and 7800GTX.

After looking at what Crossfire has to offer with both low and high end cards, we begin to see more than just potential in multi rendering solutions. We also mentioned that the performance limiting factor seems to vertex and polygon processing power. NVIDIA's latest products, the GeForce 7600 and 7900 series have all come with more vertex shader units than the previous generation. Having a much higher clock core will certainly not hurt them either. We're very interested to see whether this theory is true or not, so we put two GeForce 7600GT cards together for some SLI testing. We were pretty anxious to see if NVIDIA had a stronger offering for SLI with their GeForce 7600GT cards. Of course, it would've been more interesting to retest SLI with the high end GeForce 7900GTX, but we think both setups don't behave that much differently from one another.

At the same time, NVIDIA have also released a new reference driver, Forceware 91.31. No doubt, this latest driver is meant to support the recently launched GeForce 7950 GX2. For those who are not in the loop, the GeForce 7950 GX2 is actually two GeForce 7900 graphics cards put together into a single packaging that you can install on a single PCI Express x16 slot. The new drivers is updated with more game profiles, mostly to enable SLI support for these games. That makes it particularly interesting for our SLI testing.. However, after some time testing with these new drivers, we decided it would not be wise to use them for our performance evaluation.

Problems

Generally, we will use the more up to date drivers for an article or review. However, that is not the case here. For the time being, we will stick to Forceware 84.21 for most of our reviews. Although the new drivers offer some interesting features, like additional options for SLI multi rendering, we had a lot of issues and problems with Forceware 91.31 (and it's not the new control panel).

 

Starting with Forceware 91.31, NVIDIA have followed ATI's lead in many ways. It's not just the control panel, but also the handling of applications that made use of overlays, including FRAPS. So now, when FRAPS is active, you will not see changes to AA and AF settings from the control panel immediately applied, much like with ATI's Catalyst drivers. The safest way to enable AA and AF is to stop FRAPS first and then make the changes. Simple really. However, Forceware 91.31 have a very annoying bug regarding this behavior. The driver seems to fail purging the overlay and the screen after we exited a full screen application - like we usually do after benchmarking a game. So, after a benchmarking session, we can see the desktop, but can't interact with it. Or worse, all we see is a blank screen. There is a workaround to this: we simply have to switch to the desktop before quitting, stop FRAPS, then switch back and finally exit the game. This bug shouldn't affect ordinary users and gamers, but it really is annoying if you're trying to benchmark a game with FRAPS. Hopefully, NVIDIA will fix this in a patch or a new driver.

We also encountered other issues with Forceware 91.31. With an SLI setup, the drivers' application detection seems to be overriding control panel settings such as AA and AF for some games. Take a look at the table below - these are benchmark results with a pair of GeForce 7600GT with various modes of SLI multi rendering enabled.  The results are taken with default settings in the control panel (without AA and AF).

Call of Duty - Dawnville

	Forceware 91.31 Default
	Auto Select
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	81	63	44
Avg	129.69	88.35	62.06
Max	197	126	86
	Alternate Frame Rendering
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	80	63	44
Avg	129.58	88.14	61.97
Max	198	125	86
	Split Frame Rendering
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	80	63	44
Avg	129.2	88.28	61.95
Max	197	126	86
	Alternate Frame Rendering 2
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	80	63	44
Avg	129.55	88.3	62.06
Max	198	126	86

all numbers are in frame rates per second except for 1024 x 768, 1280 x 1024, and 1600 x 1200

Ooh, that's way to low for default settings. Now look at the results we've got from the older drivers, Forceware 84.21.

Call of Duty - Dawnville

	Forceware 84.21 Default
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	82	83	81
Avg	214.95	184.28	151.98
Max	428	332	246
	Forceware 84.21 AA AF
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	83	63	44
Avg	129.92	88.14	62.14
Max	197	125	87

all numbers are in frame rates per second except for 1024 x 768, 1280 x 1024, and 1600 x 1200

It's clear that the new drivers seems to be forcing 4x AA and 16x AF in Call of Duty, despite the fact that we left the respective AA and AF options on their default 'Application controlled' positions. There is a workaround to this, simply rename the executable to something else. Below are the results we got with the new drivers, still from Call of Duty, but this time we've renamed the Call of Duty's multi player executable file to cord.exe.

Call of Duty - Dawnville - renamed to cord.exe

	Forceware 91.31 Default
	Auto Select
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	78	80	80
Avg	210.32	191.89	165.98
Max	427	347	304
	Alternate Frame Rendering
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	79	80	78
Avg	211.28	191.58	166.17
Max	463	345	304
	Split Frame Rendering
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	79	80	78
Avg	209.92	193.11	163.88
Max	401	324	248
	Alternate Frame Rendering 2
Resolution	1024 x 768	1280 x 1024	1600 x 1200
Min	78	79	79
Avg	211.13	190.89	166.55
Max	466	345	305

all numbers are in frame rates per second except for 1024 x 768, 1280 x 1024, and 1600 x 1200

The results are right where they should be. Needless to say, this is not normal behavior. If we choose to run benchmarks with the new drivers, we have to examine the results to see if the drivers are forcing settings and overriding users' choices. Then we have to rename the games that are affected to get the real results. We think it's actually better (and a whole lot simpler) to stick with the old drivers for now, especially since there seems to be nothing to gain (performance wise) with the new drivers.

Another strange thing we've encountered with these new drivers is SLI installations with graphics cards from different vendors doesn't seem to work, at least not right away. Switching cards didn't fix the problem. It's as if the second graphics card is not properly detected. In the end, we had to fall back to the old drivers, enable SLI and then reinstall the new drivers to get SLI working. We admit, this may be just a screw up in our part, but it's unlikely.

These issues and problems are more than enough for us to say these newer drivers are not worth the upgrade (at least for us). So, for most of this article, we will be using the older, Forceware 84.21.

Performance

AA and AF settings are applied in game, whenever possible. We had to apply AA and AF from the driver panel for Call of Duty and Richard Burns Rally. In Quake 4, we enable both AA and AF from the console, with the command 'r_multisamples' set to 4 and 'image_anisotropy' set to 16. V sync was disabled both from inside the game and on the driver panel. AA and AF is enabled from the settings menu for F.E.A.R, Serious Sam II and Splinter Cell: Chaos Theory.

Call of Duty and Richard Burns Rally was configured to the highest possible detail. F.E.A.R settings are set to our usual test settings (all settings set to maximum / on except for soft shadows), but for this article, we turned off volumetric lights. Quake 4 was set to 'High Quality' while Serious Sam II was tested with our usual standard settings (maximum settings but with custom resolution, anti aliasing and anisotropic filtering settings). The same also applies to Splinter Cell Chaos Theory - the game was benchmarked to use its shader model 3.0 and everything turned on (except for HDR).

We ran our usual tests under two different configurations - one with a single GeForce 7600GT and then with two GeForce 7600GT in SLI's multi rendering mode. We've also included results with application specific profiles for comparison purposes.

We'd like to thank Tagan for supplying the additional power supply for this article.

Our test setup
AMD Athlon 64 3500+ socket 939
2 x 1024 MB Kingston KVR 3-3-3 PC3200 DDR-SDRAM
MSI K8N NForce 4 SLI motherboard
2 x GeForce 7600GT DDR3 256 MB graphics card
Maxtor DiamondMaxPlus9 80 GB Serial ATA 8 MB buffer
ASUS E-616 DVD-ROM
Tagan TG530-U15 530 watts ATX/BTX power supply

Windows XP Professional with Service Pack 2 installed
NVIDIA Forceware 84.21 reference driver
NVIDIA NForce 6.70 reference driver
Creative SoundBlaster Live! 24 bit 5.12.1.512 driver.
DirectX 9.0c

The results:

The graphs are pretty self explanatory, but in case you can't see the text, they're arranged in groups of three colors: green for minimum fps, blue for average fps and red for maximum fps. The first group represent test results under default settings, and the second represent results with 4xAA and 16xAF enabled. These results are in frame rates per second.


Overall, results from a pair of GeForce 7600GT running in SLI seems to be lower than from a single card, even when AA and AF is enabled. This is obviously not a good sign. From our past experience with a pair of GeForce 6600, Call of Duty is one of the games that really benefit from the addition of a second card. However, when we look at these results, it would seem SLI is not really offering anything significant here. Could we be system limited? We don't think so - we saw higher frame rates from a GeForce 7800GTX and 7900GTX on this very same setup. We're pretty certain we've hit the graphical limit here and from the looks of it, we're not pixel shader / texture bound.

The additional pixel processing power of the second card is mostly unused, however the additional driver overhead with SLI remains. So instead of an increase, we see a very small drop in performance. It's very small, but it's there and consistent across all resolutions. Even enabling NVIDIA's optimized profile is not going to help much. This is simply all that GeForce 7600GT SLI can offer with Call of Duty.

Compared to Crossfire, SLI is obviously not doing as good of a job with this game. If you remember, we saw an increase with a pair of Radeon X1600XTs and X1900s in default settings and with AA / AF enabled (though the increase is very small with the X1900s). But before we make any conclusions, let's look at the other benchmarks.


Well, thankfully not all games behave like Call of Duty. Surprisingly, Richard Burns Rally is one game that's extremely friendly to multi rendering solution. Even with default settings, we're already seeing the additional processing power from the second card coming into play at 1280 x 1024. With AA and AF, SLI is already showing its prowess at 1024 x 768. Once we hit 1600 x 1200, the difference is very significant indeed, both in default settings and with AA and AF. That's about 20 to 30 fps for both the minimum and average fps or about 15 percent in default settings and 60 percent with AA and AF. Those extra frames will certainly help you enjoy this game.

We think both Crossfire and SLI is on equal footing here, but only if we were to force AFR rendering with Crossfire. Obviously, SLI has the upper hand, since you can force SLI multi rendering mode from the driver panel. Those of you itching to try the new drivers will be happy to know that NVIDIA have included this game's profile in their latest driver, Forceware 91.31. No doubt it includes an SLI profile and perhaps other fixes as well. We'll take a deeper look at this later on, but for now, let's move on to the next game. 


From the looks of it, you're better off forcing multi rendering manually than using the game's SLI profile - you'll get a slightly higher frame rate in average (2 fps more). SLI can also help you achieve a high average frame rate at 1280 x 1024, but in all honesty, you probably don't want to play at that resolution. The minimum fps is way too slow - lower than 20 fps with AA and AF. Judging by these results, 1024 x 768 is the only viable choice out of the resolutions available for gameplay with this setup (even with a pair of GeForce 7600GT in SLI). At 1280 x 1024, frame rates will drop to an unplayable 20 fps.when explosions and smoke fills the screen. Certainly not a good reason to use SLI - we gain nothing here.

This particular cut scene is very hard on graphics cards. Even a high end card like the Radeon X1900XTX or the GeForce 7900GTX have trouble pushing 1280 x 1024 with AA and AF. Remember these two cards have roughly twice the bandwidth compared to the GeForce 7600GT we're using. We didn't gain anything here with SLI, but neither did we with Crossfire.


SLI definitely delivers here and rightly so. You need all the pixel processing power you can get to run this demanding game. With default settings at 1024 x 768, the SLI setup has much to offer - about 4 to 11 fps on the minimum fps. That's just enough for it to pass the 30 fps minimum mark. Gains in average fps are even more significant - about 29 fps or 65 percent faster than a single GeForce 7600GT. However, using higher resolution is not recommended, unless you're willing to lose AA and AF. If you choose to do so, you most likely will not go any higher than 1280 x 960. At 1600 x 1200, the game runs so slow, frame rates can drop to below 30 fps and that's without AA and AF.

We also saw similar gains with Crossfire, though a bit smaller in degree with two Radeon X1600XTs (6 fps on minimum, 20 fps on average) and two Radeon X1900s (20 fps on average). The GeForce 7600GT SLI setup offer higher nominal frame rates than Crossfire, but that's hardly related to the multi rendering implementation of the two manufacturers. The cards themselves are very different, so its hard to measure which solution (SLI or Crossfire) offer the higher increase. There's also other factors at play here as well.


What can we say? Frame rates are actually lower with SLI. Thankfully, it's not really noticeable in gameplay, not like what we saw with the GeForce 6600 or 7800GTX SLI. The difference is about 1 to 2 fps overall and pretty constant (except at 1024 x 768, but we think that's the extra difference here is just normal variations between runs). With the GeForce 6600 and 7800GTX, the differences are higher and more noticeable (at least on the GeForce 6600). This really is a slap in the face for SLI. With an engine that's very well suited to the GeForce architecture, SLI should really offer something much better than this.

In contrast, with Crossfire we saw an increase in performance close to 80 - 90 percent, so it's obvious who has the better support for Quake 4. The differences are unsettling at the very least. A pair of Radeon X1600XT in a Crossfire setup using AFR is actually faster than a pair of GeForce 7600GT in this benchmark - all because we gain nothing with SLI. This is a big disappointment, especially so since NVIDIA have included a profile for this game even before the game came out. Clearly, there's a lot of homework to be done if SLI can offer a challenge to Crossfire in this game.


Another disappointment. On average, we lost some frames at default settings and only gain 1 to 2 fps at most with AA and AF enabled. The trend we saw with Call of Duty, SW: KOTOR and Quake 4 continues. In this case, it would've been much wiser to buy a GeForce 7900GT or 7900GTX instead of a pair of GeForce 7600GT. It looks like we are right all along in our conclusion of SLI - you should never expect a performance increase with SLI.

With Crossfire, we saw a definite increase with a pair of Radeon X1600XT and slightly less definite increase with a pair of Radeon X1900s. Interesting to note, the increase was quite significant with the Radeon X1600XTs. So much so, that the resulting nominal average frame rates at 1024 x 768 is almost the same as what we're getting with these two GeForce 7600GTs. The Radeon X1600XTs in Crossfire mode even has a higher average frame rates at 1280 x 960 and 1600 x 1200. However, you will likely never use both setups at those resolution since the frame rates are too low for comfortable gameplay .


As always, Splinter Cell: Chaos Theory is one strong proof supporting the use of multi rendering solutions - be it Crossfire or SLI. With a pair of GeForce 7600GT, you can get away with 1280 x 1024 with AA and AF enabled. If you were only using one GeForce 7600GT (or SLI in single rendering mode), you'll have to do with 1280 x 1024 without AA and AF or 1024 x 768 with AA and AF (probably) If you look closely at the results, you'll see similar levels of frame rates between a single GeForce 7600GT at 1024 x 768 and a pair of them at 1600 x 1200. A rough calculation will indicate an increase of 70 - 80 percent in performance with the addition of a second card.

Looking Back

Let's recap what we've seen so far. We've already looked at 7 games - Call of Duty, Richard Burns Rally, SW: KOTOR, F.E.A.R, Quake 4,  Serious Sam II and Splinter Cell: Chaos Theory.  Out of these 7 games, a pair of GeForce 7600GT in SLI mode can only offer a performance increase in three games, Richard Burns Rally, F.E.A.R and Splinter Cell: Chaos Theory. Based on our previous findings, a pair of GeForce 6600 can only offer a performance increase with Call of Duty, Richard Burns Rally, F.E.A.R and Splinter Cell: Chaos Theory. We also hardly see any increase with a pair of GeForce 7800GTX, except for Richard Burns Rally and F.E.A.R. In contrast, we saw an increase in Call of Duty, Richard Burns Rally, F.E.A.R, Quake 4 and Serious Sam II with both a pair of Radeon X1600XT and a pair of Radeon X1900s. The table below clearly illustrate this.

	6600 SLI with Forceware 81.98		7800GTX SLI with Forceware 81.98		7600GT SLI with Forceware 84.21
Call of Duty
	Default	AA AF	Default	AA AF	Default	AA AF
1024	Slower	Faster	Slower	Slower	Slower	Same
1280	Same	Faster	Slower	Same	Slower	Same
1600	Same	Faster	Slower	Same	Slower	Same
Richard Burns Rally
	Default	AA AF	Default	AA AF	Default	AA AF
1024	Slower	Faster	Slower	Faster	Same	Faster
1280	Slower	Faster	Slower	Faster	Faster	Faster
1600	Slower	Faster	Slower	Faster	Faster	Faster
FEAR
	Default	AA AF	Default	AA AF	Default	AA AF
1024	Faster	Faster	Faster	Faster	Faster	Faster
1280	Same	Faster	Faster	Faster	Faster	Faster
1600	Same	Same	Faster	Faster	Faster	Faster
Quake 4
	Default	AA AF	Default	AA AF	Default	AA AF