<!–#set var="article_header" value="ATi Radeon 9700 PRO –
Pretender To The Throne” –>
Introduction
It’s like the movies. Different summer; same story. The king rules his domain with an iron grip, until a contender to the throne rises from the masses to challenge his position. If this were a movie plot, chances are the audience would scorn it, feeling cheated by yet another remake of a remake. And in reality, not even the cast has changed in this power-struggle drama: NVIDIA in its typical role as “The King,” ATi as “The Challenger.”
Yet what would indeed be a cheap attempt at cashing in on previous successes in the world of the silver screen may be no less than a case of history repeating itself in the graphics market. After all, once upon a time it was NVIDIA who was the upstart, and the now-defunct graphics pioneer 3dfx ruled 3D-land. But, irony of ironies, NVIDIA came, saw and conquered, annexing the toppled champion and integrating the remains of his empire into its own. It was also roughly at this point that the remaining competition began to be hopelessly outpaced and outclassed by NVIDIA’s developments. Meanwhile, ATi was still completely focused on the lucrative OEM market and in no way prepared to play catch-up with NVIDIA. In other words, there simply was no competition for NVIDIA.
Q4 2001 Standalone Market Share by Mercury Research. Information from Trident.
Of course, this wouldn’t be the classic story of drama and conflict if the tides hadn’t begun to change. Last year, ATi introduced its second-generation Radeon card, the 8500. Hampered at first by logistical problems during its introduction, it soon began to test the ruling champions forces, and several skirmishes ensued. Still, the fact remained that, at best, the 8500 had barely achieved parity – which was nonetheless quite an achievement. In the end, the GeForce3 proved to be superior in a number of ways. The introduction of the GeForce3 Ti line was expedited, which only served to solidify NVIDIA’s lead, giving the company some more breathing space. Only a few months later, the GeForce4 Ti followed, and King NVIDIA once again seemed unstoppable, retaining the performance crown and, thus, the upper hand in the war.
Interestingly enough, the challenger decided to switch tactics – instead of working on a refined update to the 8500 series (the way the established king would have), the Canadians concentrated their efforts on its successor. In the past months, we have witnessed all sorts of preludes to a renewed skirmish on both sides, with white-paper attacks countering driver-leak feints. Now the time has come, and ATi is launching a new offensive, lead by its newest champion, the Radeon 9700, alias R300. But enough of the drama, for now. Let’s look at the real-life impact of the new generation.
Introduction, Continued
The heart of the Radeon 9700 PRO, the “R300” in a Flip-Chip design.
Where ATi’s 8500 was able to achieve technological parity with NVIDIA’s flagship, the 9700 goes one step further, becoming the first card to fully support and implement all features specified in Microsoft’s DirectX 9 gaming API. For now, this is a victory on paper only, since the upgrade to DirectX 9 is unlikely to happen before fall of this year, probably not until the fourth quarter. As with almost every new DirectX generation, it’s hard to tell when games using these new features will even become available, although it’s safe to say that it may be a while.
Nonetheless, technological evolution is essential in this industry. Several factors work hand-in-hand here. The potential buyer needs to be convinced of the validity of the technological platform before buying a given product. The more people who buy the product, the greater the installation base, which in turn helps convince software companies to include these new features in their own product. And the more software titles that implement these features, the more interested buyers are in getting a product that also offers these features. Thus, without the technological innovation from NVIDIA, ATi and others, the state-of-the art would be stuck at the same level it was at years ago. To sum it up, although new features may not be implemented in games immediately, there is definitely a long-term payoff, in the form of feature evolution, so to speak.
Luckily, this evolution is beginning to gain momentum and speed, driven in part by Microsoft’s Xbox gaming console, which uses DirectX 8 as its 3D API. Since this makes the two platforms closely related, the PC stands to gain a lot. Consider – every game developed for the Xbox and then ported to the PC (or vice versa) must be at least DirectX 8 compliant!
In addition to offering new features and supporting new technologies, it is of utmost importance that a modern graphics card also be able to handle current standards correctly. Matrox’s new Parhelia is a good case in point. Being a “half DirectX 9 compliant part,” the card may lure many feature-conscious buyers. In reality, these features can only be seen in action in Matrox’s own demos (although admittedly to stunning effect). When used with existing software, the card disappoints and is unable to fulfill the high expectations gaming enthusiasts will have of a 3D card of this price segment. It offers nowhere near the performance of a GeForce4 Ti4600, and often even comes up short against the “old” Radeon 8500, a card that costs only a fraction as much as the Parhelia.
Manufacturers who produce graphics cards with ATi chips.
It should be noted that original ATi cards (i.e., built by ATi) will only be available in North America. In Europe, ATi will leave the distribution and sales to its established partners like Hercules, while Creative will be in charge in Asia. On top of that, many other companies like Gigabyte or PowerColor will also be introducing their own take on the 9700 theme. ATi is obviously taking its cue from NVIDIA and transforming itself from a card manufacturer into a chip-maker and seller.
In the article accompanying the product launch, we already extensively covered the new technology incorporated into ATi’s new flagship. Now, the first Radeon 9700 boards are finding their way into the retail channel and have to prove to gamers everywhere that this card is more than a white-paper wonder.
The Radeon 9700 PRO
Compared to the GeForce Ti 4600, the Radeon has pleasingly compact dimensions.
The extensive power requirement makes it necessary to have an additional cooling unit on the backside of the card.
The new chip will be available in two versions. The faster one, which will be called “Radeon 9700 PRO,” will be released first. A short while later, ATi will begin selling “normal” Radeon 9700, a version with slightly lower memory and core clock speeds. ATi seems to have learned its lesson from the confusing introduction of the Radeon 8500.
Buyers had great difficulties telling the various versions of the card apart, due to the fact that the naming convention was very misleading. It was almost impossible to tell whether you were getting the faster version or the slower OEM version (built-by-ATi vs. powered-by-ATi).
Many resellers contributed to this confusion with contradictory advertisements in their product and price guides. Now the fastest version can finally be easily identified by the “PRO” tagged onto the name. That doesn’t mean that the normal version will be stripped down in any way – it will still encompass the full set of features. The GPU – or VPU, as ATi now prefers to call it – will simply be clocked at a slightly lower speed.
The real budget version will be introduced sometime in the fall and will carry the name Radeon 9500. To save costs, it will probably be stripped of some of the 9700’s pixel pipelines and be clocked lower than the bigger cards.
The Radeon 9700 PRO, Continued
NVIDIA GeForce4 Ti4600 | ATI Radeon 8500 | ATI Radeon 9000 PRO | ATI Radeon 9700 PRO | |
Chip Technology | 256-bit | 256-bit | 256-bit | 256-bit |
Process | 0.15 Micron | 0.15 Micron | 0.15 Micron | 0.15 Micron |
Transistors | 63 Mio | 60 Mio | – | – |
Memory Bus | 128-bit DDR | 128-bit DDR | 128-bit DDR | 256-bit DDR |
Memory Bandwidth | 10.4 GB/s | 8.8 GB/s | 8.8 GB/s | 19.8 GB/s |
AGP Bus | 1x/2x/4x | 1x/2x/4x | 1x/2x/4x/8x | 1x/2x/4x/8x |
Memory | 128MB | 64/128MB | 64/128MB | 128/256MB |
GPU Clock | 300 MHz | 275 MHz | 275 MHz | 325 MHz |
Memory Clock | 325 MHz (650 DDR) | MHz 275 (550 DDR) | 275 MHz (550 DDR) | 310 MHz (620 DDR) |
Memory | BGA 2.8ns | SD/BGA 3.3ns | SD | BGA 2.x ns |
Vertex Shader | 2 | 2 | 2 | 4 |
Pixel Pipelines | 4 | 4 | 4 | 8 |
Texture Units Per Pipe | 2 | 2 | 1 | 1 |
Textures per Texture Unit | 4 | 3 | 6 | 8 |
Vertex S. Version | 1.1 | 1.1 | 1.1 | 2.0 |
Pixel S. Version | 1.3 | 1.4 | 1.4 | 2.0 |
DirectX Generation | 8.0 | 8.1 | 8.1 | 9.0 |
FSAA Modi | MultiSampling | SuperSampling | SuperSampling | MultiSampling |
Memory Optmizations | LMA II | Hyper Z II | Hyper Z II | Hyper Z III |
Display Outputs | 2 | 2 | 2 | 2 |
Chip Internal Ramdacs | 2 x 400 MHz | 2 x 400 MHz | 2 x 400 MHz | 2 x 400 MHz |
Chip External Ramdacs | – | – | – | – |
Bits per Color Channel | 8 | 8 | 8 | 10 |
Special | – | – | TV Encoder On-Chip; FullStream | TV Encoder On-Chip; FullStream |
Estimated Price | ~ $300-350 | $179 | $149 | $399 |
Compared to its competitors, the heart of the Radeon 9700, the R300 chip, can boast some very impressive technical specifications. Weighing in at 100 to 110 million transistors (ATi’s various white-papers contradict themselves in this respect) the chip sports roughly 40 million transistors more than the GeForce4 Ti. It also features four vertex shaders and eight pixel pipelines – twice as many as NVIDIA’s flagship. At least the second part of that statement is only half true, however, since the R300’s pixel pipelines can only process one texture per clock cycle, compared to the GeForce4’s two.
In real-world terms, this means that the 9700 is roughly twice as fast as the GeForce4 in single-texturing scenarios. When confronted with the much more common multi-texturing, both chips should be roughly identical in speed – at least in theory.
In practice, neither card will be able to exploit the pixel units to their full potential, since they simply can’t be fed enough data to be 100% efficient – the deciding factor here is memory bandwidth.
The Radeon 9700 PRO, Continued
A large fan cools the 0.15-micron chip with over 100 million transistors.
Thanks to its 256 bit-wide DDR interface, the R300’s memory bandwidth is almost twice as high at any given clock speed. In addition, the chip is already DDR-II capable, meaning we may see a new revision of R300-based cards using this faster memory.
ATi has also tweaked and refined the Radeon product line’s rendering optimizations with Hyper-Z III. You can find more information about what this technology actually does in the previous technology article. The advantage of an 8x AGP bus seems to be of a more theoretical nature at this point. There just aren’t any games at present that would come anywhere near needing this kind of bandwidth, much less saturating it.
The 256-bit DDR memory with a BGA design runs at 310 MHz (620 MHz DDR).
Finally, ATi has also finally addressed the previous Radeons’ greatest weakness: the extremely low performance when using the function for smoothing edges (FSAA). Instead of the slow supersampling method, the 9700 now employs the much faster multisampling technique, which is similar to NVIDIA’s implementation. Additionally, the R300’s take on multisampling is supposed to include further performance optimizations. Later on in this article, we will take a look at how fast this new method really makes the 9700.
Due to its high power consumption, the Radeon 9700 requires an external power supply, which is provided by the 3.5″ cable. The AGP slot would have been overloaded with this task.
2D Features
The R300 has the GeForce4 Ti at a slight disadvantage where 2D features are concerned. The ATi chip carries two 400MHz RAMDACs that use 10Bit precision internally. This makes the chip fully multi-monitor enabled, just like the NV25. The 165MHz TMDS transmitter also allows DVI resolutions of up to 2048×1536 (QXGA), while TV-Out works at resolutions up to 1024×768.
In order to operate two monitors in DualView mode, the Radeon 9700 offers two interfaces: VGA and DVI-I. An adapter plug (included) makes it possible to connect a second CRT monitor to the DVI-I interface. TV-out is located in the middle.
Instead of the conventional hard-coded video optimizations, the pixel shaders now take on the task of filtering video streams. ATi has dubbed this technique “FullStream.” Again, we’ll talk more about this a bit later on.
Drivers
ATi’s new Catalyst driver suite leaves a good impression, featuring many self-explanatory menus that let the user change almost any setting. Even a lot of the more advanced settings can be tweaked to a great extent. The only thing we would like to see in the control panels would be overclocking options.
The settings menu for Direct3D. The FSAA option available with the Radeon 8500, which allowed you to choose between Quality and Performance modes, is no longer offered here. For anisotropic filtering, a performance mode (bilinear) and a quality mode (trilinear) are available.
The settings menu for OpenGL offers the same options.
Detailed information about the card. The BIOS of the board is dated 2002/08/01. There’s no problem with operating an AGP 8x card in an AGP 4x motherboard.
Test Setup
And so, let the battle begin! The platform of choice is a Pentium 4 2.53GHz system from the upper performance segment,, since any user willing to spend $300 for a 3D card will more than likely also have a current system with a fast processor and a generous helping of RAM. For users with a less state-of-the-art system, we also conducted some tests on a slightly slower Pentium 4 1.6GHz system and a PIII 800 machine to show the card would perform on their own system.
We picked NVIDIA’s GeForce4 Ti4600 as the reference card for performance testing. Although Matrox’s Parhelia can be found in the same price segment, it doesn’t offer sufficient 3D performance to qualify as a contender to the performance crown. Read more in the article Attack Out Of The Blind Spot: Matrox Parhelia-512.
Hardware Main Test System | |
Processor | Intel Pentium 4 2,53 GHz (133 MHz) |
Memory | 2 x 256 MB, PC 333, CL2 |
Mainboard | ABIT BG7 (i845G) |
Hardware CPU Scaling – Moderate Performance | |
Processor | Intel Pentium 4 1,6 GHz (100 MHz) |
Memory | 2 x 128 MB RDRAM |
Mainboard | ASUS P4T (i850) |
Hardware CPU Scaling – Low Performance | |
Processor | Intel Pentium 3 800MHz |
Memory | 2 x 128 MB, PC 133, CL2 |
Mainboard | ASUS CUSL2 (i815E) |
Drivers & Software | |
Graphics Driver | NVIDIA – v30.82 ATI – v6.13.10.6143 |
DirectX Version | 8.1 |
OS | Windows XP Professional |
Benchmarks & Settings | |
Giants Retail | Version v1.04 |
Aquanox | Retail Version v1.17 |
Dungeon Siege | Retail Version v1.00 |
Comanche 4 | Benchmark Demo V 1.0.0.1.18 |
Quake III Arena | Retail Version 1.17 Benchmark using ‘Q3DEMO1’ |
Jedi Knight II | Benchmark using ‘jk2ffa’ |
3D Mark 2001 SE | Pro Version, Build 330 |
Detail Benchmarks | PowerVR Villagemark v1.17 NVIDIA Chamelon Mark NVIDIA BenMark Matrox SharkMark 3D Mark 2001 SE |
The suite of tests can be divided into several types. First, we have the typical gaming benchmarks that test performance while a game is in progress. Next come the synthetic benchmarks that rate gaming performance (3DMark2001 SE), followed by theoretical benchmarks that focus on specific areas like triangle count and fill rate. Lastly, we test image quality by looking at anisotropic filtering and FSAA.
A word on overclocking – the current version of Powerstrip already offers overclocking options for the Radeon 9700 PRO. Since the test samples are all from a pre-production batch, we decided against overclocking tests, as these would not necessarily reflect the potential of the final production boards. Consequently, we will wait until we have samples of final shipping ATi cards (or those produced by ATi’s partners) before testing their overclocking potential. The final cards to be shipped will be produced in China.
Real World – Games and Synthetic Benchmarks
Giants
Giants is a typical DirectX 7 game that uses only moderately sized textures. The game engine supports hardware T&L and requires an incredible amount of rendering power, even compared to modern games.
The Radeon 9700 Pro leaves the GeForce4 Ti trailing behind. The gap only widens with increasing resolution.
Aquanox
Aquanox is based on DirectX 8 technology. The power of a card’s vertex shader determines its performance in this game.
Again, NVIDIA’s flagship is overwhelmed by the sheer rendering power of the Radeon 9700 PRO. In 1280×1024, the GeForce4 already drops below the magical 60fps barrier, while the Radeon can almost maintain this minimum up until 1600×1200.
Dungeon Siege
Dungeon Siege is an RPG based on a T&L compatible DirectX 7 engine.
Dungeon Siege’s 3D engine has an automatic detection routine to determine the video card in use and then set the detail levels accordingly. If a card is not recognized, the engine defaults to the most basic low-quality settings. ATi supplied us with an updated configuration file for the game, which adds the Radeon 9000 and Radeon 9700 to the list of supported cards. NVIDIA’s GeForce4 was already supported. Of course, both cards use different default settings, but when the GeForce4’s settings were changed to match those of the Radeon, it resulted in slightly lower performance, so we decided to let each card run with its own default setting and call it fair.
The results are nothing short of amazing! While the performance of the GeForce4 drops with increasing resolution, the Radeon 9700 is able maintain the same high level of performance in all resolutions.
Comanche 4
Comanche 4, a DirectX 8 game, employs vertex and pixel shader effects. However, it is very CPU-dependent, meaning that, even with the fastest available CPU, these cards would hardly begin to break out into a sweat. The same does not hold true for slower video cards, however. This makes the game more of a test of overall system performance than a video benchmark. Since many other games share this characteristic, it is nonetheless important to have such a game in the test suite. Imagine buying a brand new 3D card to speed up your supposedly slow system only to find out that the graphics card wasn’t the bottleneck at all. All that money spent for nothing. That’s exactly why we run these tests.
In the “lower” resolutions, the GeForce4 Ti is able to eke out its first lead, albeit a tiny one.
Quake 3
The game that lent its engine to so many others may not be the freshest, but that doesn’t seem to diminish its popularity. Its game engine can be found in several modern games in refined, modified and tweaked versions that exploit its potential even more. Nonetheless, this OpenGL game is still THE standard benchmark in the hardware community to determine a 3D card’s pure performance, since it scales very well with increasing rendering performance.
The Radeon 9700 PRO breaks the 300 fps mark in 1024×768 – at High Quality settings and with maximum texture quality, no less! The ATi card holds on to its clear lead throughout all tested resolutions.
Jedi Knight II
Jedi Knight II is based on a modernized Quake 3 OpenGL engine, although a heavily modified and optimized one. Additionally, it employs high-resolution textures and much more complex 3D models.
In 1024×768, both cards are tied for first place. The game proves to be very CPU dependent. The higher the resolution becomes, the clearer the lead from the ATi card. Even the performance drop at 1600×1200 is almost negligible.
3DMark 2001 SE
No round of testing would ever be complete without the requisite 3DMark 2001SE run. Despite the controversy surrounding this synthetic benchmark among our readership, we nonetheless consider its final score a very good way of judging a card’s 3D performance.
The Radeon’s lead is, in a word, impressive; the generation gap between it and the GeForce4 obvious.
3D Performance in Detail
A card’s performance in a game is influenced by many factors. If a game uses large textures, then fill rate is important. If it uses complex 3D models, triangle count becomes the crucial feature. Lastly, pixel and vertex shader effects are also very demanding. The following detailed tests allow us to observe the Radeon 9700 Pro and the GeForce4 Ti4600 in a side-by-side comparison.
Triangle Count
The polygon count of 3D models, i.e. the number of triangles used to create them, is continually increasing in modern games. Some examples of this are the greater detail in the cars of a racing game, or large outdoor levels with lots of vegetation, etc.
3DMark 2001
This test has the Radeon 9700 clearly leading the GeForce4 Ti, both with extra light sources enabled (eight light test, lower score table) and in the simple version. The Radeon’s perfectly round scores seemed strange, though. It achieved the same score across all resolutions, while the GeForce4 showed a certain spread. We have yet to find an explanation for this behavior.
BenMark 5
BenMark is a small benchmark program developed by NVIDIA to test the T&L performance of a card.
In BenMark, the Radeon attains scores three times as high as those of the GeForce4 – despite the fact that it has “only” twice as many vertex shader units. This leads us to conclude that the Radeon’s individual shader units are more efficient than those of the GeForce4 at almost identical clockspeeds.
This test measures the amount of data that is transferred over the bus. Again, the Radeon 9700 can take a lot more strain than its competitor.
The Radeon is the clear winner in the category “highest triangle count.” The GeForce4 Ti simply can’t keep up.
Fillrate
The fillrate tells us how fast the pixel pipelines can process simple textures. Here are the detailed 3DMark2001 SE results:
This value shows the performance when using only a single texture. The advantage of the Radeon’s eight pixel pipelines becomes immediately apparent here. NVIDIA’s GeForce4, on the other hand, only has four pixel pipes, and is therefore outclassed.
The multitexturing test (two textures per object, for example, when using light maps) shows that ATi’s “eight pipe pixel engine” also has disadvantages, as each pipe can only process a single texture per clock cycle. The GeForce4’s pipelines can process two textures each, and thus the performance of the two cards is almost identical. Unfortunately for ATi, most current games almost exclusively use multitexturing environments.
The fill rate category leaves us with a mixed conclusion. The advantage to having eight pixel pipelines is not as great as ATi’s marketing machine would have us believe, especially in multitexturing environments. Then again, this comment is only valid with respect to fill rate. As the following pixel shader tests will show, eight pipelines can make a lot of sense.
Pixel Shader Tests
Now let’s compare the pixel shader performance of these two cards:
The Nature scene is one of the game benchmarks. The Radeon 9700 is almost twice as fast as the Ti4600.
The Radeon continues to dominate the pixel shader tests completely.
NVIDIA ChameleonMark
The last pixel shader test in our repertoire comes from NVIDIA. ChameleonMark was created to demonstrate the pixel shader performance of the GeForce3 line.
The results are, once again, amazing. The Radeon achieves three times the performance of the GeForce4 board.
ATi’s eight pipe architecture really gets to flex its muscles here. The pixel shader performance is practically twice that of the GeForce4, which is hardly surprising, considering the Radeon has twice as many pixel pipelines as the NVIDIA board.
Vertex Shader Tests
Vertex shaders are responsible for calculating the geometry of a scene. The Radeon 9700 has four vertex shaders; again, twice as many as the NVIDIA’s GeForce4.
3D Mark 2001 SE 330
As expected, the results show the Radeon securely in the lead.
Matrox SharkMark
SharkMark is a benchmark developed by Matrox, which was published for the introduction of the new Parhelia cards. The test runs in windowed mode with a predefined resolution.
Once again, the Radeon 9700 takes the lead.
The Radeon can clearly be said to dominate the vertex shader tests. However, apparently the individual vertex shader units are no faster than their counterparts in the NVIDIA design.
Occlusion Culling
Small words (kind of), big effect. Occlusion culling is a feature that optimizes the rendering of 3D scenes. If an object in a scene is blocked by another object, it would be a great waste of GPU-time (and bandwidth) to nonetheless render this blocked object. Besides needlessly increasing the triangle count of a scene, the (invisible) object would also need to be textured or have pixel shader effects applied to it. To reduce this waste (overdraw), both ATi and NVIDIA have developed techniques to detect these hidden objects and exclude them from the rendering process.
PowerVR Villagemark
With its Villagemark, PowerVR created a benchmark that tests the performance of these culling algorithms. Originally, it was meant to showcase the advantages of the PowerVR tiling architecture of the Kyro line of cards.
The results speak clearly. Even though the Radeon 9700 has a higher triangle throughput to begin with, we can conclude that ATi’s Hyper-Z implementation does its job well.
FSAA Performance
Following NVIDIA’s lead, ATi has now also switched to the faster multisampling edge-smoothing technique. You can find more details on this technology here. Let’s take a look at how the two contestants perform under Direct3D and OpenGL with FSAA enabled. To give you a more detailed impression of the overall performance, we will also include the final 3DMark 2001 SE score with FSAA.
3D Mark 2001 SE 330 – FSAA
The Radeon obviously benefits from its superior memory bandwidth. Even with 2X FSAA, it is still faster than the GeForce4 without FSAA! With 2x or 4x FSAA enabled, it offers almost twice the performance of the GeForce4 at corresponding levels.
3D Mark 2001 SE 330 – FSAA – Dragothic
In Dragothic, the Radeon once again proves its superiority.
Quake 3 – FSAA
Again, the Radeon almost reaches the performance levels of a GeForce4 Ti without FSAA. When both cards are in 4xFSAA mode, the ATi is almost twice as fast as the NVIDIA board.
In FSAA, ATi’s Radeon easily shows the competition who’s boss. Its FSAA performance is practically twice that of the GeForce4 Ti4600 across the board. Even with 2X FSAA enabled, it is usually faster than the NVIDIA board without FSAA.
Anisotropic Filtering
Anisotropic filtering is a technique that makes textures look cleaner and crisper on objects when viewed at a very small angle. More details can be found here. Radeon 9700 allows for anisotropic filtering with either a bilinear or a trilinear filter. NVIDIA, on the other hand, offers only the trilinear filtering mode, which produces better quality with the drawback of being more performance-intensive. Both cards were set to filtering level 8x.
ATi’s implementation uses a couple of tricks to lighten the burden on the GPU and improve performance. For example, the level of filtering used is automatically adjusted depending on the angle at which an object is viewed. This is a pretty smart idea, since lower (and, thus, faster) filtering levels are often sufficient at smaller viewing angles. Of course, NVIDIA has a different take on this, and the debate on which implementation is better is currently in full swing. At the moment, the only way to tell any difference between the two methods is to take screenshots and go over them with a magnifying glass. (It also helps if you squint. Oh well.) In a real-world environment while a game is in progress, it is unlikely that a user would be able to tell the difference. Once again, the Radeon’s good results are also due in large part to the eight pixel pipelines.
3D Mark 2001 SE 330 – Anisotropic
ATi’s anisotropic filtering implementation proves to be superior, thanks to the eight pixel pipelines, of course.
3D Mark 2001 SE 330 – Anisotropic – Dragothic
The detailed results of the Dragothic test show how much of a nosedive the GeForce4’s performance takes when anisotropic filtering is enabled. In comparison, the performance hit for enabling anisotropic filtering is marginal on the Radeon.
Quake 3 – Anisotropic
The Radeon also proved to be superior in OpenGL. However, ATi’s current Catalyst driver has a bug. Changing the quality of the filter from bilinear to trilinear (or vice versa) in the driver menu had absolutely no effect. During testing, we had to set the filtering level by hand, by changing a key in the Windows registry.
ATi’s implementation of anisotropic filtering clearly offers superior performance. While NVIDIA can truthfully say that their scores were achieved using a more “honest” method, this doesn’t really change anything if the results of these two methods are practically indistinguishable from one another. Admittedly, there were one or two times when ATi’s algorithm had trouble selecting the appropriate filtering level because of an especially awkward viewing angle – but even then the better performance more than outweighed that slight and momentary inconvenience. As a last resort, you could also still crank the filtering settings up all the way to 16x aniso and still get a higher frame rate than with NVIDIA’s honest method.
Maybe ATi should simply include a complete but unoptimized implementation of anisotropic filtering in its driver (i.e. one that is set by hand and doesn’t automatically select the “appropriate” filtering level). That way, everybody would be happy.
Maximum Quality – Aniso + FSAA
This is a torture test that will bring even the highest of the high-end cards to their knees in the high resolutions. The goal is to attain the highest possible quality by enabling both anisotropic filtering and edge smoothing at the same time.
3D Mark 2001 SE – FSAA + Aniso
The results show the GeForce4 at a dramatic disadvantage. At a resolution of 1600×1200, the Radeon 9700 is almost three times faster.
3D Mark 2001 SE – FSAA + Aniso – Dragothic High
The performance ratio remains the same in the Dragothic test.
Quake 3 – FSAA + Aniso
In Quake 3, the Radeon 9700 is “only” twice as fast as the GeForce4 Ti.
In this mode of testing, the GeForce4 Ti loses dramatically to the Radeon 9700. It is interesting to note that the performance delta between the cards is much more pronounced in Direct3D than it is in OpenGL. In light of these results, NVIDIA may have to reconsider its anisotropic filtering implementation and its position in regards to honesty. We’ll say the same thing here that we said about ATi’s implementation, though – in the end, the user should be free to decide whether or not an optimization should be enabled.
CPU Scaling
Many benchmark results look very impressive but do little to help the prospective consumer reach an informed decision. New graphics cards should be tested on the fastest available hardware to demonstrate their full potential. Yet how many buyers actually have the same hardware at home that hardware testers have at their disposal? Reading a review such as this, many may find themselves wondering which subsystem really needs to be upgraded the most. CPU? Mainboard? Video card?
To help you make an informed decision, we have also conducted a few tests on slower systems, namely a Pentium 4 1600MHz and a Pentium 3 800MHz system.
Keep in mind, though, that we can only make very broad statements here. If a game is mostly CPU-dependent, then even the fastest video card won’t be able to speed up your system. Conversely, if a game requires a lot of rendering power, then a high-end card may make sense even in a slower system – especially if you’re planning to use FSAA and/ or anisotropic filtering.
Here is a list of several games, each with its own requirements:
Aquanox
Aquanox’s 3D engine is quite independent of CPU power. It does rely very heavily on the video card’s rendering performance, though. At 1280×1024, a P3 system produced almost the same results as a P4 2.53! This phenomenon can be observed on both the GeForce4 Ti and the Radeon 9700 PRO.
Quake 3
Quake 3 is completely different in this respect. This game just loves CPU power. While the Radeon and GeForce4 returned almost identical results on a P4 1600MHz, it took the jump to the P4 2.53GHz to unleash the Radeon’s full potential and let it stretch its legs.
3DMark 2001 SE
The 3DMark 2001 SE score is influenced both by the CPU and the video card – which makes sense, since it was designed as a system benchmark.
So, a fast CPU doesn’t necessarily guarantee a better end result. Or the other way around: a faster video card won’t necessarily give you higher frame rates – at least not in low resolutions. In plain English: if you’re considering buying a Radeon 9700, you should ensure that the rest of the system can keep up with it.
FullStream Video
Back in the Radeon 9700 technology article, we gave you a quick preview of ATi’s new FullStream video technology. The Radeon 9700 is no longer equipped with hard-wired video acceleration features. The pixel shader has now taken on this task. In addition to the usual video filters, this allows for real-time effects and optimizations of the video stream.
One possible practical application would be Internet video stream, whose images are often blocky due to heavy compression and low resolution. FullStream is supposed to be able to compensate for these defects and smooth the image, increasing overall image quality.
Contrary to ATi’s original claims, applications do need to be adapted to make use of the FullStream technology and its features. Standard videos do not profit from these new techniques. So far, the only application that is “FullStream ready” is a special version of the RealPlayer. According to ATi, a FullStream patch for the Windows Media Player is already in the works and is currently undergoing beta-testing. It is unknown at this point whether the DivX platform will also benefit from this technology in the future.
Unfortunately, we can’t give you an image quality comparison between the standard RealPlayer and its optimized sibling, since the latter version did not allow us to take any screenshots, making a side-by-side comparison impossible. Even specialized screen-capture programs like Hypersnap were unable to create screenshots, since the program would need direct access to the hardware – which it obviously did not get. Instead, let me offer my subjective impression of this feature.
Although the amount of blockiness and the number of artifacts were indeed reduced, this came at the cost of overall image clarity and crispness, which caused some of the details in the image to become unrecognizable. Fortunately, this feature can be disabled in the RealPlayer.
FullStream left us with a decidedly mixed impression. As a feature, it’s “nice to have,” but not quite mature enough to be a serious selling point.
Conclusion
The King is dead! Long live the King! How’s this for a plot-twist? The challenger Radeon – a real “Performeron” – has actually done it and usurped the throne from the former king! ATi has earned itself not only the performance crown in gaming environments, but also that of the technology leader!
The Radeon 9700 PRO proved to be superior in all possible categories, be it the framerate while the game was in progress, the triangle throughput, FSAA, anisotropic filtering, or pixel and vertex shader performance. NVIDIA’s flagship trails the new champion in every discipline. And Matrox? In light of these results, the Parhelia looks even less attractive than before.
To put these results in perspective, keep in mind that the GeForce4 Ti is already half a year old. Then again, it is also still selling at $300 to $350, even today. The recommended retail price of the Radeon 9700 has been quoted as being $399. We expect to see retail prices even lower than that in computer hardware stores. Besides, the “normal” 9700 will be even less expensive. As a result, we should see the price of GeForce4 Ti boards drop very rapidly. This will present NVIDIA with another challenge: to readjust prices across the entire GeForce4 family (both Ti and MX) and justify the existence of some of its members. Of course, this all depends on ATi’s ability to ship their boards in sufficient quantities – which is exactly where the Radeon 8500 failed.
Where 2D features are concerned, neither card offers a feature that would make it more desirable than its competitor. Both the R9700 and the GeForce4 Ti offer dual-monitor support, including a very comfortable management software as well as TV-Out functionality. However, the Radeon has a slight advantage due to its 10Bit DACs and the higher DVI resolution.
If you’re looking for the fastest gaming card around, then you’re looking for the Radeon 9700. NVIDIA’s counteroffensive in the form of the highly anticipated NV3x won’t be more than a rumor for at least a few more months. Currently, we expect cards to be on store shelves for the holiday season. As for the update to the GeForce4 family, code-named NV18 and NV28, these cards are still a few weeks off, and it is doubtful whether a simple update with higher frequencies would be sufficient to come anywhere near the 9700’s performance levels. Even if it did (a big “if”), these parts would still lack the Radeon’s DirectX 9 support.
Of course, NVIDIA will undoubtedly answer this challenge, and we are more than just a little curious as to the tricks that the Californians have up their sleeves which would help them reclaim the throne that they’ve occupied for the past years. ATi is also making preparations, though, and the next performance update in the form of DDR-II memory is already on track for the end of this year. The pared-down Radeon 9500 will put additional pressure on NVIDIA in the mid to low range market. On the whole, ATi’s chances for success don’t look bad at all. And the real winners of this performance revolution are the consumers. After all, the most important factor in deciding which product to buy is still price, not performance. Thanks to the ongoing and bloody feud between the big boys of computer graphics, we can expect prices to continue dropping. Whoever said it was right: “A revolution every now and then is a very healthy thing.”