ATI’s Radeon 9800 XT graphics card

The Radeon 9800 XT is based on ATI’s R360 chip, which is essentially a minor revision of the R350 chip used in the Radeon 9800 Pro. The chip is still fabbed by TMSC using 0.15-micron process technology. ATI has modified R360 silicon in order to clean up some trouble areas and to allow for higher clock speeds. Beyond that, the R360 has all the familiar R350 features, including support for floating-point pixel formats, programmable shaders, and F-buffer logic that enables longer shader programs without conventional multipass rendering. With the tweaks, the Radeon 9800 XT has reached new highs: a 412MHz core clock speed with an effective 730MHz clock speed on its DDR memory. That’s just a step or two up from the 380/700MHz core and memory speeds of the 9800 Pro 256MB.

Other changes from the Radeon 9800 Pro to the XT are similarly simple. The 9800 XT card is a new circuit board design, visibly distinct from the 9800 Pro 256MB. Also, to keep the R360 chip cool at its higher clock speeds, the Radeon 9800 XT comes stock with a fancy new copper cooler, like so:

The Radeon 9800 XT sports a fancy new spoiler.. er, cooler

The rear copper plate acts as a heatsink, too

This new cooler is larger and heavier than the almost-wimpy heatsink/fan combos on previous Radeon 9800 cards, but it’s nothing compared to the massive hunks of metal strapped to every GeForce FX 5900 Ultra. The larger fan on the XT’s cooler varies its speed according to chip temperatures. It runs a lower speeds during normal use, and kicks up to a faster speed when the R360 chip gets a workout. Running full tilt, the XT’s cooler isn’t horribly loud, but it’s definitely audible. When it’s rotating slowly, the fan is nearly silent. Of course, the same could be said for the cooler on the Radeon 9800 XT’s primary competitor; the GeForce FX 5900 Ultra’s blower is quiet at low speeds, and not too terribly loud going flat out.

With the new cooler, new PCB layout, and tweaked silicon, ATI says the 9800 XT is more overclockable than its forerunner, although we haven’t had the card long enough to test that theory just yet.

As you can see in the pictures, the 9800 XT comes with the standard array of outputs: one VGA, one DVI, and one video out. We’ve long wished to see these top-end cards come with dual DVI outputs and a pair of adapters for VGA, but apparently neither ATI nor NVIDIA has gotten that message yet. Hope springs eternal here at TR, though, so we’ll keep mentioning it.

The clock speed equation
The higher clock speeds of the Radeon 9800 XT mean better performance, particularly in the pixel-pushing department. From simple pixel filling to complex programmable pixel shader effects, the primary bottlenecks to all-out performance in graphics today are related to pixel pushing. The Radeon 9800 XT is all about high resolutions and a healthy dose of (both edge and texture) antialiasing in current games, plus fluid frame rates in next-generation games. To get a sense for what the 9800 XT brings to the table, let’s bust out one of our world-famous chip charts.

With another 32MHz over the Radeon 9800 Pro, the 9800 XT theoretically improves on its predecessor only a tad, but the 9800 Pro was already a behemoth. I won’t go into all the dynamics of it, because you can read the chart yourself, and because we’ll be testing things here in a moment.

Note how the Radeon 9800 XT matches up against the GeForce FX 5900 Ultra. With eight pixel pipes, the XT blows it away in pixel fill rate. The 5900 Ultra is faster with multitexturing, however, and has even more memory bandwidth than the XT. In these simple, theoretical measures, it’s a tight race.

The cooler won’t occupy an adjacent PCI slot

Also coming soon: the Radeon 9600 XT and CATALYST: Overdrive
ATI will also be introducing an XT version of its mid-range Radeon 9600 product soon. The new RV360 chip, on which the 9600 XT will be based, is a tweaked version of the chip on the Radeon 9600 Pro. It’s essentially a four-pipeline version of the R300 VPU from the original Radeon 9700. Like the current Radeon 9600 Pro, the Radeon 9600 XT will be fabbed with TSMC’s 0.13-micron process. Unlike the current 9600, though, TSMC will use a low-k dielectric in fabricating the new RV360 chips, which should reduce the RV360’s power requirements and allow for higher clock speeds. As a result, the Radeon 9600 XT should hit core clock speeds upwards of 500MHz, and ATI will pair it up with memory running at over 600MHz.

ATI says the Radeon 9600 XT will have a small, quiet cooler, and it won’t require a secondary power connector like the 9800 XT. 9600 XT cards will cost about $199 when they first arrive some time in October, and to sweeten the deal, all new Radeon cards will soon be shipping with Half-Life 2 in the box.

In November, ATI plans to unleash CATALYST: Overdrive, new software with dynamic overclocking capabilities built in. ATI says this is an “intelligent software solution that determines the optimal ASIC speed for maximum performance.” Without crashing, of course. ATI also says the increased clock speeds will be “ATI Quality Assured,” so you won’t void your card’s warranty by playing with this feature. As I understand it, only newer Radeon chips with built-in temperature sensors will allow for dynamic overclocking.

I expect real overclockers will find better performance by overclocking their cards manually, but CATALYST: Overdrive should allow casual overclockers a chance at some additional performance for free—and that’s what overclocking is all about. Personally, I can’t wait to give this feature a shot with a well-cooled Radeon 9600 XT.

Some frank words about our testing
Because of the recent controversies over various graphics benchmarks and game benchmarks, we are somewhat unsure how to handle graphics performance testing properly right now. Both ATI and NVIDIA have cheated on benchmarks in the past, but NVIDIA has been on the warpath against any benchmark that would show its GeForce FX chips in a negative light from the time of that product line’s first release. We have seen new NVIDIA driver revisions consistently use a range of tricks, from less-rigorous filtering to outright replacement of shader code, to improve frame rates at the expense of image quality.

We would like to take the time, when a new product like this one is released, to conduct detailed testing with innovative benchmarking methods and exhaustive image quality comparisons. ATI recommends as much to reviewers, as well. However, our Radeon 9800 XT review sample showed up on our doorstep this past Friday morning, so we were not able to spend as much time as we’d have liked testing the 9800 XT against its competitors. As a result, we’ve had to restrict our testing to a mix of old and new games, plus a couple of interesting new graphics benchmarks.

To better understand all the issues involved in graphics evaluation today, I spent some time last week talking with NVIDIA’s Chief Scientist, David Kirk. Mr. Kirk is an intriguing guy, part engineer and all politician, very much capable of telling you exactly what he wants you to hear and nothing more. During the course of our conversation, I took away a couple of important bits of information that are relevant here.

As I questioned him about why the GeForce FX cards seem to perform rather poorly in games with lots of DirectX 9 pixel shaders, Mr. Kirk repeatedly told me not to fixate on issues of color precision and datatypes. He admitted the NV3x chips are very “sensitive” to optimizations, but explained that color precision isn’t the issue so much as instruction ordering is. He explained that a better compiler in the GeForce FX driver, translating DirectX API calls into NV3x instructions, could vastly improve performance, and that just such improvements are in store in the 50-series drivers from NVIDIA. When I asked him if he was confident that NVIDIA could wring good performance from the NV3x chips in general cases, without the need for application-specific optimizations, he answered affirmatively, although he conspicuously stopped short of swearing off the need for app-specific work.

I left the conversation with the distinct impression that I needed to test NVIDIA’s 50-series drivers. As a result, I’ve tested the GeForce FX 5900 Ultra with NVIDIA’s 51.75 drivers, which are at present a press-only release. I prefer to test with public release drivers, but I’ve made an exception here.

Also, I’ve included results for several benchmarks that we know are somehow compromised by application-specific optimizations that reduce image quality for performance. Both ATI and NVIDIA seem to have taken steps to reduce the texture filtering loads on their chips in Unreal Tournament 2003. NVIDIA’s legal team apparently scared FutureMark into allowing optimizations in 3DMark03, including wholesale shader replacements, that FutureMark originally disallowed. Other publications have raised questions about image quality compromises in newer NVIDIA drivers when running AquaMark3. The list goes on. Please know we are aware of these issues, and we have included benchmark results because we had to test something. At the very least, you can compare the 9800 XT to the 9800 Pro 256MB, knowing the two are most likely going about their tasks in entirely similar ways in a given test.

Because of the short time we’ve had with Radeon 9800 XT, we’ve limited our testing to situations where $499 video cards excel: situations with high fill-rate demands, lots of edge and texture AA, and lots of pixel shaders. In many cases, that means we’ve limited our testing to 1600×1200 resolution in 32-bit color, with and without 4X edge antialiasing and 8X anisotropic texture filtering. In other cases, with newer games or more complex scenes, lower resolutions sufficed.

Our testing methods
As ever, we did our best to deliver clean benchmark numbers. Tests were run at least twice, and the results were averaged.

Our test system was configured like so:

We used ATI’s CATALYST 3.7 drivers (version 7.93) on the ATI cards, which is a full release driver that’s been through Microsoft’s WHQL certification. The NVIDIA 51.75 drivers are a beta release not available to the public.

Thanks to Corsair for providing us with memory for our testing. If you’re looking to tweak out your system to the max and maybe overclock it a little, Corsair’s RAM is definitely worth considering.

The test systems’ Windows desktops were set at 1152×864 in 32-bit color at an 85Hz screen refresh rate. Vertical refresh sync (vsync) was disabled for all tests.

We used the following versions of our test applications:

• FutureMark 3DMark03 Build 330
• Quake III Arena v1.31 with trdemo1.dm_67
• Wolfenstein: Enemy Territory with demo0000.dm_82
• Serious Sam SE v1.07 with Demo0003
• Unreal Tournament 2003 with trtest1.dem
• Splinter Cell v1.2 with TRKalinatekDemo.bin
• X2 rolling demo
• Gun Metal benchmark v1.20
• ShaderMark 2.0
• AquaMark3
• rthdribl 1.2

All the tests we employed are publicly available and reproducible. If you have questions about our methods, hit our forums to talk with us about them.

Benchmark results
We’ll start with some older games and move toward some newer games and benchmarks as we go.

Quake III Arena

Serious Sam SE

The Radeon 9800 XT is, as expected, consistently just a bit faster than the Radeon 9800 Pro 256MB. However, in these older games, the GeForce FX 5900 Ultra is generally faster still. Only with edge and texture AA cranked up in Q3A does the Radeon 9800 XT outrun the 5900 Ultra.

Unreal Tournament 2003

Wolfenstein: Enemy Territory

These games are a bit newer, and the Radeon 9800 XT performs relatively better in them. The 5900 Ultra is fastest in Wolfenstein: Enemy Territory, but only with edge and texture AA disabled.

Splinter Cell

This Xbox port runs best on NVIDIA hardware, but the 9800 XT closes the gap a bit. When we look at the frame rates graphed over time, the cards perform very similarly overall. The 5900 Ultra does have one performance dip, just a little ways into the scene, that the ATI cards don’t experience.

3DMark03

With all the controversy over 3DMark, I won’t say too much about these results. Make of them what you will. However, the synthetic fill rate tests do seem to be showing us about what we’d expect there. The Radeon cards have a big advantage in single-textured fill rate, but the GeForce FX 5900 Ultra has a similar advantage in multi-textured fill rate.

AquaMark

The Radeon 9800 XT’s additional MHz just moves ATI ahead of NVIDIA in AquaMark3. With 4X antialiasing and 8X anisotropic filtering, however, even the 9800 Pro pulls past the 5900 Ultra.

Gun Metal

X2 Demo

Gun Metal and the X2 demo are based on newer game engines. Like Splinter Cell, Gun Metal is an Xbox port that runs well on NVIDIA hardware. The X2 demo is a little odd. It has some great visuals, but the game gets choppy at high resolutions instead of slowing down gracefully as it approaches fill-rate limits.

Real-time high-dynamic-range lighting effects

This nifty little app shows off the promise of DirectX 9 about as well as anything I’ve seen. Unfortunately, GeForce FX cards don’t seem to run it quite right. On FX cards, some colors are “off,” and math errors are visible in the lighting effects. Also, FX cards produce visible banding in color gradients.

HDR lighting on my lowly Radeon 9500 Pro..

…and on the GeForce FX 5900 Ultra with 51.75 drivers The screenshots above are but one example. Here is an uncompressed version of the Radeon 9500 Pro screenie, and here is the uncompressed version of the same scene rendered by a GeForce FX 5900 Ultra. Perhaps this is just a case of an application incompatibility, but if not, it seems that differences in color precision are not always difficult to spot.

Incidentally, the Radeon 9800 XT shows a nice performance gain over the Radeon 9800 Pro in this application.

ShaderMark 2.0
This new version of ShaderMark is an interesting animal. It’s a DirectX 9 benchmark that makes extensive use of Pixel Shader 2.0, and its various shaders are written in Microsoft’s High-Level Shading Language. The registered version of ShaderMark 2.0, which we’re using, can provide “partial precision” hints to tell NV3x cards to use 16 bits of precision per color channel in pixel shader calculations instead of 32. (ATI R3x0 chips always use 24 bits per color channel.) This program also includes some special “2.0+” shaders with extra features for the NV3x GPUs.

Also, because this is a brand-new version of ShaderMark, nobody’s driver is optimized for it yet, so we shouldn’t be seeing any funny business with shader replacements or the like. This version does include an “anti-detect” mode to defeat app-specific driver optimizations, but we found it didn’t change scores.

Some of the shaders would not run on the GeForce FX 5900 Ultra, especially the more complex ones that make multiple passes. As I understand it, part of the problem is the GeForce FX’s lack of support for floating-point texture formats. Anyhow, you’ll see some blank scores below as a result.

When it can complete the tests, the 5900 Ultra is much slower than the Radeon 9800 Pro and 9800 XT. Partial precision hinting helps in many cases, but not enough to allow the 5900 Ultra to reach more than about 60% of the 9800 XT’s performance. This performance delta between GeForce FX and Radeon cards is reminiscent of what we saw in Half-Life 2 not long ago. Even with the 51.75 drivers, the general pixel shader performance of the GeForce FX 5900 Ultra doesn’t look good.

The 9800 XT, by contrast, just improves on a good thing.

ATI Radeon 9800 XT September 2003

The same can’t be said for NVIDIA’s GeForce FX products, and that fact causes us significant concern about future game compatibility and performance. Even the new 51.75 drivers from NVIDIA, with an updated compiler, couldn’t make the GeForce FX 5900 Ultra competitive with ATI products in ShaderMark or the high-dynamic-range demo we tested. One gets the impression NVIDIA needs a well-stocked sweatshop full of driver software developers working 24×7 in order to keep the GeForce FX competitive in various games and benchmarks as they are released. ATI’s hardware, by contrast, just seems to work right. Whatever the reason for this disparity, ATI cards seem to provide a much smoother user experience, especially for power users and cutting-edge gamers who are always downloading the latest and greatest games and demos.

Because of these concerns, and because the Radeon 9800 XT delivers kick-ass overall performance, we are doing something we don’t always do around here: busting out an Editor’s Choice award. The Radeon 9800 XT does well on its own merits, but in this case, we also want folks to know there’s a clearly superior choice between the two $499 graphic card exotics. Five hundred bucks is a lot of cash to lay out on an AGP card, so making the right call is important if you plan to do so. No, the 9800 XT isn’t wildly faster than the Radeon 9800 Pro 256MB that it replaces, but it does perform better at the same price. These cards aren’t good values by any stretch—buy a Radeon 9600 Pro if it’s value you want—but the Radeon 9800 XT is our graphics exotic of choice.