instantwebmeetings.com - Video Conference, Collaboration, E Learning, Video Collaboration, Video Meeting, Unified Communications

Introduction

High-end video card launches tend to receive the lion’s share of the media coverage, but sales of $450-$700 behemoths aren’t where the money’s at in the video card industry. For all the focus on the 3870 X2 and 8800 GT Ultra in two-, three-, four-, or 64-way SLI configurations, both NVIDIA and ATI sell the bulk of their product and make the majority of their profits at a sub-$200 price point. The 9600GT we’re reviewing today is NVIDIA’s latest flagship in the $170-$200 price range.

NVIDIA’s midrange product line was formerly anchored by the 8600 GTS, but that card has become increasingly unattractive over the past several months. ATI’s HD 3850 decisively outperformed the 8600 GTS at the same price when it launched last November, while the HD 3870 and 8800GT slotted themselves nicely into the $200-$300 market. With the launch of the 9600GT, NVIDIA plans to retake the midrange market and leave the HD 3850 eating dust.

G94: The card, and its corresponding bells and whistles

The 9600GT has some impressive stats for a mid-range card, as indicated by the chart below. Architecturally speaking, the G94 sits much closer to the G92 than the G84 sat to the G80. While G84 was a more efficient architecture than G80, the gap between the two cards was quite significant. With just 1/3 of the G80’s streaming processors and 40 percent of its bus width, the 8600 GTS’s performance always lagged the 8800GTS 320MB’s significantly.

The 9600GT, on the other hand, is made of sterner stuff. It uses the same 256-bit memory bus as the 8800 GT and the 8800 GTS 512, and carries twice the number of SP’s that the G84 does. All of these factors should make the performance gap between the 9600 GT and 8800 GT smaller than the gap that existed between the G80-based GPUs—if the card delivers as promised, it should be a potent midrange competitor indeed.

One other feature of the 9600 GT that’s worth noting is the card’s increased 512MB RAM loadout. Neither ATI nor NVIDIA defines a single amount of memory that’s “right” for any particular card, but all recent midrange products from both companies have used 256MB of RAM as part of the reference design. NVIDIA’s decision to step up to 512MB, in this case, isn’t just a marketing ploy to win marketshare—there are games on the market today that’ve outgrown the 256MB limitation and crater when forced to run at high detail levels with such a small amount of memory.

Although the card we’re testing today offers a standard set of two dual-link DVI ports, the card is capable of other options. All 9600 GT cards support HDCP and HDMI output via either a DVI adapter or through a native port if the manufacturer provides one, and DisplayPort is natively supported with the 9600 GT. NVIDIA has also included a few new goodies in its PureVideo HD implementation, but we’ll examine those separately later in the review.

Does the 9600 GT have what it takes to challenge the HD 3850’s sweet price/performance ratio? Let’s take a loo.

System configuration, test results

The HD 3870 and HD 3850 card(s) used in this review are stock models provided by ATI. The 8800 GT’s are Alpha Dog editions from XFX. Alpha Dogs are clocked at a 670MHz core / 975MHz memory and are some of the fastest (if not the fastest) 8800GTs available today. XFX was also kind enough to provide the 8600 GTS we used to obtain the results covered below.

Testbed configuration was as follows:

• Acer AL2216WBD 22″ LCD
• Intel Core 2 Duo E6850 (3GHz)
• Asus P5N-E SLI (NVIDIA 650i chipset)
• 4GB Crucial Ballistix DDR2-1067 memory
• PC Power and Cooling 750W PSU
• Windows Vista Ultimate (32-bit)
• ForceWare release 169.25 (WHQL)
• ForceWare release 173 (Beta 9600GT driver)
• Catalyst driver 8.2 (ATI)

3DMark 2006 Although it’s a synthetic benchmark, 3DMark 2006 is easily accessible and generally familiar to anyone with even a passing interest in measuring 3D performance. The results here shouldn’t be read as indicative of performance in any particular game, but should generally indicate how numbers will trend.

At default settings, the 9600GT only maintains a 7 percent lead over the 3850, but the gap widens nearly 15 percent once we crank the detail levels up. The 8800GT is about 25 percent faster than the 9600GT in HQ mode, and the HD 3870 beats NVIDIA’s mid-range darling by a comfortable 18 percent. The 8600GTS is beaten by every card we tested, and by whopping margins in all cases. The 9600GT is no less than 83 percent faster than the 9600GTS it replaces, and these type of synthetic results don’t bode well for the 8600GTS’s performance in actual games.

World in Conflict

World in Conflict is an RTS game developed by Massive Entertainment and published by Sierra Entertainment earlier this year. The game uses DirectX 9 for rendering when detail levels are set to “Medium” and switches to DirectX 10 when we change to HQ settings. Although we’re using it as a video card benchmark, WiC performance also varies depending on CPU type and speed, making it an all-around useful performance test.

ATI scores well in World in Conflict at our normal quality settings. The HD 3870 wins this round of competition over the 8800GT, while the HD 3850 beats out the 9600GT. Turning up the detail settings (and switching to DX10 mode) flip-flops the results. The 8800GT and 9600GT place first and second in HQ mode, followed by the HD 3870, 3850, and the 8600GTS. Pay particular attention to the performance gap between the two ATI cards in HQ mode. The 3870 is clocked faster than the HD 3850, but the gap between the two video cards’ scores can’t be explained by clock differentials alone. At least part of what we’re seeing here is the need for 512MB of video RAM when running in high quality modes. As for the 8600GTS, it manages a playable framerate at medium quality levels, but tanks in HQ mode—even compared to the 3850, which is equipped with a similarly-anemic amount of RAM.

Crysis, Company of Heroes

Crysis is a benchmark that scarcely needs introduction—the engine on which the game is built has already proven itself capable of bringing even the mightiest of gaming machines to its knees. This may not have done much to help the game’s disappointing sales, but it has established Crysis as the gauntlet that all high-end video cards have to run.

Crysis runs quite well at medium detail levels. NVIDIA cards have historically held an advantage in Crysis, and the 9600GT proves true to this record. All of the cards in question maintain playable framerates, though the 8600GTS is still choppy at times. Relative performance rankings remain identical when we shift to high quality testing, but the performance hit is quite significant. Of the various cards we tested, only the 8800GT and 9600GT maintain semi-playable framerates. As for the 8600 GTS, again, there’s not much to say when a card’s performance could almost be charted in seconds-per-frame instead of frames-per-second.

The 9600 GT returned a markedly higher minimum framerate than the 8800 GT, even when repeatedly tested. I suspect this may be the result of further driver optimizations.

Company of Heroes

Company of Heroes is another RTS game and was originally released as a DX9 title. The game has since been patched to include DX10, and it allows the end user to customize virtually every graphics option you can think of—which should, in theory, make it an excellent benchmark.

In this case, I’ve tested both Medium and HQ detail levels in DX10, using the game’s built-in benchmark. Before moving on to the graph, though, there’s a point I need to address regarding the 8800GT’s minimal framerate. While testing the card, there were a few points where the game would hiccup, for lack of a better word. When this occurred, framerates would drop dramatically for less than a second.

All the cards maintain a playable framerate at average quality settings, though the 8600GTS lags the rest of the pack. In HQ mode the HD 3870 sets the pace, but both the 8800 GT and the 9600 GT perform well. Again, we see the 256MB cards stagger at HQ settings compared to their 512MB brethren, though the 8600 GTS and the HD 3850 end up nearly tied for last place in this particular test.

HD decoding

NVIDIA is using the launch of the 9600 GT to reveal additional PureVideo technology that was baked into the G92 but kept hidden up to this point. Once the ForceWare 174 drivers become available, users with 8800 GT, 8800 GTS, or 9600 GT cards will be able to use several new features that weren’t previously exposed, including dynamic contrast enhancement, dynamic color enhancement, and dual-stream decode. All three are fairly self-explanatory.

Dynamic contrast enhancement, when enabled, applies what NVIDIA refers to as an optimal correction curve to frames with non-optimal exposure. According to the company’s literature, this is “akin to applying Photoshop’s ‘Auto Levels’ function up to 30 times a second.”

Dynamic color enhancement is a new feature intended to punch up blue, green, and skin tone colors. If the feature works correctly, characters and objects should look more vibrant without simultaneously seeming over-saturated. NVIDIA provided the following image as an example of how dynamic color enhancement is supposed to work, though based on my experience, it’s definitely a “best case” scenario.

I’ve included test results for the 9600 GT with dynamic color and contrast turned off, and with them turned on. HD decoding was tested in the movies Serenity, which uses the VC-1 codec, and Transformers, which uses AVC. CPU usage was monitored using Vista’s included Performance Monitor. Samples were taken once per second over a period of ten minutes. In Transformers, I began measuring at chapter 2, as the US Army is on patrol over the desert. In Serenity, I began testing at the start of chapter 3, as the Firefly-class vessel begins planetfall. I also spot-checked various points in both movies to ensure that my segment choices represented CPU load accurately.

MPEG-2 decoding performance was tested using a 1080i clip provided by a fellow Ars staffer. The clip runs some 30 seconds, was played back directly from the hard drive, and was looped over a period of ten minutes to achieve an accurate measurement of average playback.

HD movie playback was handled by the Xbox 360 HD DVD player, which connects to the testbed via USB2 and played back via CyberLink PowerDVD 7.3. For the time being I’ll continue to evaluate HD performance using HD DVD, though a transition to Blu-ray is undoubtedly in the cards at some point. The HD 3850’s decode performance can be considered representative of both the HD 3870 and the HD 3870X2—I’ve tested all three cards and have seen no significant variations in performance or quality.

The first thing I want to note is that video quality appeared virtually identical across all the cards I tested. The HD 3850 has an overall lead in VC-1 processing, as ATI’s UVD handles this step completely in hardware, while NVDIA’s PureVideo does not. In H.264 the CPU usage gap between ATI and NVIDIA was much smaller, as both architectures handle H.264 completely in hardware. The 8600 GTS and 8800 GTS log nearly identical CPU usage in both codecs, but the 9600 GT used notably more processing power. This is probably the result of immature drivers. Enabling NVIDIA’s two new “dynamic” features is very nearly free in both VC-1 and H.264.

MPEG-2 decoding favors NVIDIA, though the total amount of CPU time required for decode is small in all cases. Again we see the 9600 GT requiring a bit more oomph to get the job done as compared to the G86/G92 processors, and again, the cost of enabling the card’s new PureVideo capabilities is quite small. The true lesson to take away from these numbers is that any of the mid-range (or above) cards from ATI and NVIDIA are capable of handling modern video decoding and handle it quite well. I didn’t have time to test NVIDIA’s dual decode technology, but intend to follow up on that in the near future.

I’m not sure what to make of NVIDIA’s new dynamic contrast and dynamic color technologies. I think I noticed a slight difference in my MPEG-2 clip, but Serenity looked identical. If contrast or color were adjusted in that movie, the differences were too subtle for my eyes to catch. Transformers, on the other hand, did look different in some spots, but the differences weren’t good. Enabling dynamic color enhancement in that film created odd patches of orange on the soldiers’ faces at the very beginning of the movie.

I suspect that both dynamic color enhancement and dynamic contrast enhancement are best viewed as tools for improving low-quality video rather than as options one would leave on all the time. Viewed from that perspective, both are welcome features that could come in quite handy under specific circumstances.

Power consumption

All power consumption figures reflect total system power consumption as obtained at the wall using a Kill-a-Watt meter.

In absolute terms, none of the video cards we tested today are particularly power-hungry. The 8800 GT is notably more efficient then the 8800 GTS 320 it replaced, while the HD 3850 and HD 3870 cards were an immense improvement over the 2900XT. With that said, however, the 9600 GT sets a new record for performance-per-watt efficiency. The 9600 GT draws slightly less power than the HD 3850 at idle, but manages to outperform ATI’s mid-range card while drawing 30W less when both systems are put under load. As for the 8600 GTS, the numbers say it all. Not content with stomping all over the ATI card’s performance, the 9600 GT has delivered a coup de grâce by decisively beating the 8600 GTS’s power consumption at the same time.

Conclusion

Simply put, the 9600 GT is one hell of a card. Less than a year after it launched the 8600 GTS, NVIDIA has created a card that surpasses its predecessor in every way. The 9600 GT’s 512MB of RAM and impressive performance could make it the must-have option in the $169-$189 space—assuming, of course, that vendor prices actually end up matching NVIDIA’s suggested MSRP. This could actually take a few weeks, depending on initial supply and demand, but once card pricing comes down, ATI will have a serious fight on its hands. The company knows it, too—HD 3850 and 3870 prices have been quietly dropping at Newegg. The 256MB HD 3850 is now available for as little as $154.99, with a 512MB flavor at just $169.99 and an HD 3870 as low as $184.99. ATI’s HD 3000 series should remain competitive at these price points, but it’s a safe bet that the company would’ve preferred to keep its higher margins.

If we were still giving out numerical scores in reviews, I’d say the 9600 GT deserves its “9,” both literally and figuratively. It crushes the 8600 GTS in virtually every game benchmark available and is a much, much, better part then an “8650″ or even an “8700″ moniker might have indicated. Historically, NVIDIA’s numerical jumps do not always equal the appearance of an entirely new architecture. The GeForce 3, GeForceFX, GeForce 6, and GeForce 8 all introduced significant new features, while the GeForce 2, GeForce 4, and GeForce 7 are more properly characterized as refined and enhanced versions of the architectures that immediately preceded them. Based on what we’ve seen thus far from the 9600 GT, I expect the higher-end entries in the series to deliver stronger performance without introducing a fundamentally new architecture.

The only real losers in all of this are the buyers who recently invested in an 8600 GTS. In all fairness, however, the writing has been on the wall for that particular card since the HD 3850, 3870, and 8800 GT collectively pounded it last November. Going forward I expect that the HD 3850 and 3870 will remain price/performance competitive, thanks to the price cuts AMD is meting out, but I’m not sure what ATI will do or can do at this point to address the gap between the 3870 at $185 and the 3870X2 at $449. As price gaps go, that one’s a monster, and simply cutting the price of the 3870X2 really isn’t going to do the trick.

SOURCE