
Ah - Spring. Mystical time of rebirth. Nature's starting point in the cycle of the seasons. And, not to forget, that wondrous time of year when graphics companies, both Canadian and American, put their new chips on the table. And once again, the poker game for the biggest market share is in full swing. You could say that NVIDIA has already "played its cards" with the introduction of its new product line-up - the high-end FX 5800, the mid-range FX 5600, and the entry-level FX 5200. That means it's ATi's turn. Now ATi is playing its own trump card in the mid-range segment in the form of the Radeon 9600 and 9600 PRO, which should be available through the retail channel come May and are aimed squarely at NVIDIA's FX 5600 and FX 5600 Ultra.
Offering a very attractive price/ performance ratio, this market segment is the most interesting for the consumer. Aside from the low-cost OEM market, this is where the card makers must try hardest to gain and retain their slice of the pie. High-end cards, on the other hand, are more a matter of prestige than profit. The low quantities in which they are produced simply don't offer enough of a margin to get by on.

The Radeon 9600 PRO reference card.
The Radeon 9600, also known as the RV350, is taking the place of the 9500 series in ATi's product line, which was based on the R300 (Radeon 9700). With the introduction of the Radeon 9500 series, ATi created a good deal of confusion, as the two chips of that family differ in the number of pixel pipelines that they offer. While the 9500 PRO and all cards of the 9700 series feature eight pipes with one texturing unit each, that number was reduced to four in the non-PRO version, adversely affecting gaming performances, especially when using anisotropic filtering. The Radeon 9500 PRO, meanwhile, was head and shoulders above the competition, namely NVIDIA's GF4 Ti 4200, thanks to its fast FSAA and anisotropic filtering performance. While the DirectX 9 support definitely didn't hurt either, it remains a theoretical advantage to date, due to a lack of titles for that API.

The backside of the card.
ATi was never able to turn much of a profit with the Radeon 9500, since it is only "artificially" throttled. Quickly, adventurous modders found ways to re-enable the four deactivated pipelines. The trouble is that the 9500/ 9500 PRO chips have the same amount of transistors as the high-end models, weighing in at a hefty 107 million transistors. That's quite a lot for a mid-range chip. Compared to the relatively moderate 63 million transistors of a GF4 Ti 4200, that makes the 9500 a very expensive chip indeed.
In the end, ATi was able to beat NVIDIA's competing products - but, quite literally, at a price. The end result was a dearth of chips, since ATi was unable (or unwilling?) to keep up with demand. The chip simply cost ATi too much for the market at which it was aimed.

Photo of a wafer with NVIDIA GeForce FX 5600 GPUs.
The new Radeon 9600/ 9600 PRO remedies this situation. Like the Radeon 9500, the RV350 chip has only four pixel pipelines, but physically, this time. That spells the end of the popular mods witnessed with the 9500 - these are no longer physically possible. The changes go even further, though. The number of vertex shader units has also been slashed from four to two. This brings the chip's transistor count down to a much "healthier" (think bottom line here) 75 million transistors. Combined with its 0.13µ production process, the production cost for the chip is now down to the regions of its direct rival, the GeForce FX 5600 with its roughly 80 million transistors. This should enable ATi to rejoin the game for market share with lower stakes.
Feature-wise, the RV350 is very similar to the R300. The main difference is the new chip's 0.13µ production process, which allows for higher clock frequencies while at the same time lowering production cost. The RV350 also inherited SmoothVision 2.1 with its improved "color compression," as well as "Hyper Z III+" with an optimized Z cache and improved stencil operations from the new R350 (Radeon 9800). The pixel and vertex shader capabilities are identical to those of the R300 generation.

BGA memory with 3.1ns.
The RV 350's memory bus is 128 bits wide, just like the Radeon 9500's. What separates these two chips is the reduction of pixel pipelines from eight to four (compared to the Radeon 9500 PRO), and of the vertex shaders - from four to two. Our benchmark section should be able to tell us to what extent ATi can make up for these disadvantages by increasing the clock speed (400 MHz, up from 275 MHz). Talking about pixel pipelines, we shouldn't forget that numbers aren't everything - these pipes need to be kept "fed." All the pixel pipelines in the world wouldn't do a chip any good if the memory bandwidth was insufficient to keep them supplied with data.

- Radeon 9600 PRO 128 MB DDR (400/600); available: April 2003; price: $169-$199;
- Radeon 9600 64/128 MB DDR (325/400); available: April 2003; price: $149-$169.
And the features of the Radeon 9600 VPU (RV350) in summary:
- DirectX 9;
- ~75 million transistors;
- Four pixel pipelines (4x1 design);
- Two vertex shader units;
- 0.13 micron manufacturing process;
- 128 bit DDR Memory (DDR-II ready);
- Up to 128 MB Memory;
- SmartShader 2.0;
- SmoothVision 2.1 (optimized FSAA, AA and memory controller; 6:1 color compression);
- HyperZ III+ (8:1 Z-compression).
| ATI Radeon 9500 | ATI Radeon 9500 PRO | ATI Radeon 9600 PRO | NVIDIA GeForce 4 Ti 4200-8x | NVIDIA GeForce FX 5600 Ultra | |
|---|---|---|---|---|---|
| Chip Technology | 256 bit | 256 bit | 256 bit | 256 bit | 256 bit |
| Process | 0.15 Micron | 0.15 Micron | 0.13 Micron | 0.15 Micron | 0.13 Micron |
| Transistors | ~107 Mio | ~107 Mio | ~75 Mio | ~63 Mio | ~80 Mio |
| Memory Bus | 128 bit DDR | 128 bit DDR | 128 bit DDR | 128 bit DDR | 128 bit DDR |
| Memory Bandwidth | 8.8 GB/s | 8.8 GB/s | 9.6 GB/s | 8.0 GB/s | 11.2 GB/s |
| Pixel Fillrate | 1.1 Gigapixel/s | 2.2 Gigapixel/s | 1.6 Gigapixel/s | 1.1 Gigapixel/s | 1.4 Gigapixel/s |
| Anti Aliased Fillrate | 6.6 Billion AA Samples/s | 13.2 Billion AA Samples/s | 9.6 Billion AA Samples/s | 4 Billion AA Samples/s | 5.6 Billion AA Samples/s |
| AGP Bus | 1x/2x/4x/8x | 1x/2x/4x/8x | 1x/2x/4x/8x | 1x/2x/4x/8x | 1x/2x/4x/8x |
| Memory | 64/128 MB | 128 MB | 128 MB | 64/128 MB | 128 MB |
| GPU Clock | 275 MHz | 275 MHz | 400 MHz | 250 MHz | 350 MHz |
| Memory Clock | 270 MHz (540 DDR) | 275 MHz (550 DDR) | 300 MHz (600 DDR) | 257 MHz (514 DDR) | 350 MHz (700 DDR) |
| Memory | -unknown | BGA 3.6 ns | -unknown | BGA 3.1 ns | -unknown |
| Vertex Shader | 4 | 4 | 2 | 2 | 1 (Array) |
| Pixel Pipelines | 4 | 8 | 4 | 4 | 4 |
| Texture Units Per Pipe | 1 | 1 | 1 | 2 | 1 |
| Textures per Texture Unit | 16 | 16 | 16 | 4 | 16 |
| Vertex S. Version | 2.0 | 2.0 | 2.0 | 1.1 | 2.0 |
| Pixel S. Version | 2.0 | 2.0 | 2.0 | 1.3 | 2.0 |
| DirectX Generation | 9.0 | 9.0 | 9.0 | 8.1 | 9.0 |
| FSAA Modi | MultiSampling | MultiSampling | MultiSampling | MultiSampling | MultiSampling |
| Memory Optmizations | Hyper Z III | Hyper Z III | Hyper Z III+ | LMA II | IntelliSample |
| Optmizations | SmartShader 2.0
SmoothVision 2.0 |
SmartShader 2.0
SmoothVision 2.0 |
SmartShader 2.0
SmoothVision 2.1 |
nFiniteFX II
Accuview |
CineFX
IntelliSample |
| Display Outputs | 2 | 2 | 2 | 2 | 2 |
| Chip Internal Ramdacs | 2 x 400 MHz | 2 x 400 MHz | 2 x 400 MHz | 2 x 400 MHz | 2 x 400 MHz |
| Chip External Ramdacs | - | - | - | - | - |
| Bits per Color Channel | 10 | 10 | 10 | 8 | 10 |
| Special | TV Encoder On-Chip; FullStream
Adaptive Filtering |
TV Encoder On-Chip; FullStream
Adaptive Filtering |
TV Encoder On-Chip; FullStream
Adaptive Filtering |
TV Encoder On-Chip;Adaptive Filtering, DirectX 9+ | |
| Estimated Price | $179 (old price) | $199 (old price) | $169-$199 | $169-$199 | $169-$199 |
In addition to the raw performance of a card, image quality is becoming more of a deciding factor for the prospective buyer. Edge smoothing, or FSAA (full scene anti-aliasing), and texture filtering through anisotropic filtering tremendously improve the visual quality of games. Since these features are also the biggest performance killers, both ATi and NVIDIA are investing a lot of effort into developing optimizations that achieve a good balance between quality and performance.

The ATi Direct3D.
Using "tricks" to improve performance is most common in anisotropic filtering, to such an extent that one could argue the feature no longer deserves the name. ATi even calls it "adaptive anisotropic filtering," meaning not all textures are filtered at the level set in the driver. Instead, the driver or the chip determines the level of filtering needed based on the position and distance of a texture. This setting is controlled through the options "performance" and "quality" in the driver menu.
Obviously, the best results are achieved when both features are used in combination. The Radeon 9500/ 9600/ 9700/ 9800 family offers a full 36 different combinations: standard mode, three FSAA modes, and eight anisotropic modes (quality/ performance), for a total of 24 FSAA/ anisotropic settings alone! If that seems like a lot to you, take a look at NVIDIA's GeForce FX 5600/ 5800: standard, 42 FSAA modes (balanced, quality, performance, plus a "texture sharpening" setting for each), and nine levels of anisotropic filtering. That makes for a grand total of 378 individual FSAA/ anisotropic combinations and 429 possibilities overall!!!

For comparsion: NVIDIA's 3D quality settings.
Good luck keeping all of those straight! We decided to limit our tests to "quality" mode using the most important (comparable) modes (2x/ 4x FSAA, 8x anisotropic filtering and 4x FSAA + 8x anisotropic) so we would have enough time to bring you more than just a list of benchmark results.
We're still waiting for both NVIDIA and ATi to release official and WHQL-certified drivers before we make any final judgments on image quality. A first comparison of FSAA and anisotropic filtering quality can be found in our Radeon 9800 review here . Below are some new image comparison screen shots with anisotropic filtering. Click on the image to view the screen shot in its original size. We recommend a resolution of 1024 x 768!
Jedi Knight II (OpenGL)
Click for uncompressed and unscaled BMP image.
Take a look at the wall in the foreground and upper part of the wall at the top. You can see that the NV "performance" setting is nowhere near as clear as ATi's. Even in "quality" mode, the front wall is still fuzzy compared to Ati, and it only becomes sharp in "application." In the ATi "performance" setting, you can see a little unsharp area right of the upper torch. It's almost gone in the quality setting. NV's "quality" setting has the same difficulty with this area, which looks just fine in "application."
Unreal Tournament 2003 (D3D)
Click for uncompressed and unscaled BMP image.
In the lower part of the image, you can see the difference between performance and quality with the NV card. The white specks on the floor are slightly fuzzy in NV aniso, but nice and sharp in ATi "performance."
Unreal Tournament 2003 (D3D)
Click for uncompressed and unscaled BMP image.
You can see the same difference on the sloping wall. ATi "performance" looks much sharper. ATi "quality" also looks a bit sharper than NV "quality." Zoom in and take a look at the black areas in the texture.
My conclusion is that ATi "performance" looks much better than NV "performance." ATi "quality" also looks better than NV "quality" (just take a look at the wall in jk2), but the difference is less pronounced. Lastly: ATI also offers superior quality with 16x. This helps in certain types of games like racing simulations with large viewing areas (as you can see in the GP4 screen shots I posted in our R9800 review .
These tests were conducted on a Radeon 9700/ 9800 and a GeForce FX 5800. The Radeon 9600/ 9600 PRO, sharing the same IQ features, should yield identical results. Likewise, the quality of the FX 5600 should be identical to that of the FX 5800.
All tests were run on an Athlon XP 2700+ system built around an ASUS A7N8X (nForce 2) motherboard. In addition to the standard in-game benchmarks, we also tested the cards' FSAA and anisotropic filtering performance in Unreal Tournament 2003 (Direct3D) and Serious Sam: The Second Encounter (OpenGL).

To give you a better context for the scores, we also included the following cards:
NVIDIA: GeForce 4 Ti 4200-8x (128 MB), FX 5200 Ultra, FX 5600 Ultra;
ATi: Radeon 9000 PRO (128 MB), 9500 (128 MB), 9500 PRO, 9700 PRO.
Since both ATi and NVIDIA have published new drivers in the last few weeks, the scores from our FX 5200/ FX 5600 preview are no longer up-to-date. If you would like to compare the scores presented here with those of an FX 5800, Radeon 9800, GF4 Ti 4800 or GF4 MX 440-8x, feel free to check out said preview.

Our attempt to underclock the Radeon 9600 PRO to 9600 speeds unfortunately resulted in massive image corruption. The scores we attained seemed questionable at best, which is why they are not included here.
Using reference cards for overclocking tests is problematic: due to the normal variation in production quality, the results are never representative, especially not for the final boards of the retail card makers. Therefore, these results shouldn't be taken as definitive. The good news is that we were able to overclock the Radeon 9600 PRO. Whether this will work for the retail cards or if ATi or a card maker decides to "clock-lock" the chips remains to be seen.

The reference card can be overclocked.
The most interesting comparison in this group is between the GeForce FX 5600 Ultra, the Radeon 9600 PRO, and the Radeon 9500 PRO. The GeForce 4 Ti 4200 can be viewed as an inexpensive alternative. The Radeon 9700 PRO represents a step up into the next price category.
| Hardware | |
|---|---|
| Test System | AMD Athlon XP 2700+
ASUS A7N8X Deluxe (nForce 2) |
| Memory | 2 x 256 MB, PC 333 (2/2/2/5) |
| Drivers & Software | |
| Graphics Driver | NVIDIA - v. 43.45
ATI - v. 3.2 (6.14.01.6307) |
| DirectX Version on System | 9.0 |
| OS | Windows XP Professional SP1 |
| Benchmarks & Settings | |
| Aquanox 2 Revelation | Retail Version v2.159 |
| UT 2003 Full | Retail Version v2166 |
| Splinter Cell | US Demo 2 - Beta Bench |
| Serious Sam: Second Encounter | Retail Version v1.07
(Demo: Valley of the Jaguar) |
| 3D Mark 2001 SE | Pro Version, Build 330 |
| 3D Mark 2003 | Pro Version |
| Codecreatures | PRO Benchmark v1.0.0 |
Unreal Tournament 2003: Standard Test

Let's begin with Unreal Tournament 2003, a Direct3D game. Originally, we planned to publish minimum frame rates for this game, as well. We were forced to exclude those scores since they varied too greatly between benchmark runs however, which made it impossible to obtain clear results. The Radeon 9000 PRO still crashes in the UT FSAA/ aniso tests with the newest drivers.



The Radeon 9600 PRO can't live up to the standards set by its predecessor, the 9500 PRO. The FX 5600 Ultra also overtakes it in this test. As we already saw in the FX 5200/ FX 5600 test, the GF4 Ti 4200 scores remarkably well here.



At first, the 9600 PRO holds its own against the 9500 PRO, but falls behind with increasing resolution. The FX 5600 takes the performance lead here.



The same holds true for 4xFSAA. In higher resolutions, the 9500 PRO outperforms the 9600 PRO, which dukes it out with the FX 5600 Ultra for third place. The "old" Ti 4200 can't keep up with any of these cards.



The pared-down pixel pipelines become painfully apparent when anisotropic filtering is used. Thanks to its higher clock speed, the 9600 PRO is able to stay ahead of the 9500, but it doesn't stand a chance against the 9500 PRO with its eight pipelines.



In this combination test, the R9600 PRO leads the FX 5600 Ultra in the "playable" low resolutions. However, the R9500 PRO remains the better choice.




The cards' scores were much closer in this game. It is interesting to note that the R9600 PRO is able to beat the 9500 PRO at lower resolutions - a result of its higher core and memory clock speeds. Once again, the Ti 4200 offers remarkably good results.



No surprises here.



Close scores once again. The Ti 4200's good performance is especially surprising, considering its outdated FSAA implementation.



At first, the 9600 PRO is neck and neck with the R9500 PRO in 4xFSAA. Its performance drops off with increasing screen resolution, also placing it behind the FX 5600 Ultra.



Serious Sam shows us a different picture than UT 2003. Here we see the 9600 PRO and 9500 PRO very close. The 9600 clearly leads the FX 5600 Ultra.



The 9500 PRO once again takes the combination category (within its class). Still, the 9600 PRO bests the FX 5600 Ultra in the lower resolutions.

The Splinter Cell tests were run at the highest quality settings the game allows.



In Splinter Cell, the Radeon 9600 PRO performs very well, clearly beating the 9500 cards. The FX 5600 Ultra, on the other hand, scores very badly here. The Ti 4200's performance is once again a welcome surprise.



The minimum fps results paint a similar picture.




Aquanox 2 Revelation was also tested with the highest possible quality settings. The greatest surprise in this test is the miserable performance of the FX 5600 Ultra. The reason for its uncharacteristically poor performance remains unclear. Unfortunately, it's impossible to determine when a card uses DX 8.1 shaders versus the DX 9 version.



In this benchmark, the 9600 PRO comes in just ahead of the 9500, while the 9500 PRO with its eight pixel pipelines is clearly superior. The FX 5600 trails the group. The real surprise is once again the comparatively good performance of the Ti 4200.



The results are similar in this test.


For the sake of completeness we are also including the composite 3DMark score achieved at the standard 1024 resolution.

The Radeon 9600 PRO has a hard time in this test and clearly gets bested by the FX 5600 Ultra, and even by the Ti 4200.
3D Mark 2001: Fillrate Single Texturing

The results are slightly confusing. By rights, the Radeon 9500 PRO should score much higher. Maybe this eight-pipe card is suffering from insufficient memory bandwidth here, with the result that the pipes aren't filled with data quickly enough. FX 5600 Ultra and Radeon 9600 PRO are practically tied.
3D Mark 2001: Fillrate Multi Texturing

The scores of the 9600 PRO took us by surprise here - it seems that the higher core clock speed is partly able to make up for the lack of pixel pipelines. The FX 5600 Ultra doesn't fare as well in this test.

In this test, the "missing" two vertex shaders really hurt the 9600 PRO, resulting in lower polygon throughput.
3D Mark 2001: Vertex Shader Speed

Interestingly, the lower polygon performance of the 9600 PRO compared to the 9500 is not as pronounced in the 3DMark 2001 vertex shader test. The fact that the vanilla 9500 scores higher than the PRO version can be attributed to our review sample's higher clock speeds.
3D Mark 2001: Pixel Shader Speed

The 9600 PRO offers solid performance here. It can almost close the gap to the 9500 PRO, and clearly outpaces the FX 5600.
3D Mark 2001: Adv. Pixel Shader Speed

It remains unclear why both FX cards fall this far behind in this test, since they both support PS 1.4 in hardware, thanks to DX9 support. It's worth noting that the 9600 PRO once again trails the 9500 PRO by a wide margin.


Again, we're including the overall 3DMark score in the standard resolution.
3D Mark 2003: Game 2

Battle for Proxycon sees the Radeon 9500 PRO just ahead of the 9600 PRO and the FX 5600 Ultra.
3D Mark 2003: Game 4

The FX 5600 Ultra doesn't seem to like "Mother Nature" (Game 4).

While they didn't seem to influence the vertex shader scores in 3DMark 2001 SE, the 9600 PRO's missing vertex shaders clearly manifest themselves in the scores here. Nonetheless, the 9600 PRO is still almost twice as fast as the FX 5600 Ultra!
3D Mark 2003: Pixel Shader Speed 2.0

The pixel shader portion of the suite again sees the Radeon 9600 PRO outperform the FX 5600 Ultra.

Newer is not automatically synonymous with better, as this review shows. From ATi's perspective, the Radeon 9600 PRO is a success, as it is able to beat the FX 5600 Ultra in the majority of the tests. At the same time, it costs less to produce than the older 9500 PRO model. From the consumer's perspective, the newcomer is a step backwards, though, considering the better overall performance of the "old" 9500 PRO. The performance gap between the generations becomes especially apparent when FSAA and/ or anisotropic filtering is used. Since ATi's main concern is selling cards and making a profit, their step, backwards or not, is understandable.
Compared to the FX 5600 Ultra, the 9600 PRO seems the better choice. For the most part, it offers higher performance and also takes the lead in image quality, thanks to excellent FSAA and anisotropic filtering. The FX 5600 Ultra also seems to have trouble with some modern games like Aquanox 2 (with DX 9 shaders) and Splinter Cell. The results of the 3DMark 2003 subtests were also disappointing. Then again, the possibility remains that NVIDIA will be able to remedy this with a newer driver revision.
Despite its good performance, the gap between the 9600 PRO and the high-end cards remains huge, as the direct comparison with the 9700 PRO shows, not to mention the 9800 PRO. The FX 5600 Ultra suffers from the same problem. On the other hand, this should make high-end buyers happy, since they are really getting premium performance for the premium they're paying. It `s been a while since twice the price could be equated with twice the performance, but here it holds true, at least where FSAA and anisotropic filtering are concerned.
If you don't need the very best in image quality enhancements, can do without DirectX 9 and are happy with 2x FSAA, you'll be glad to hear that there's a very affordable card out there for you - the GeForce 4 Ti 4200. Despite its age, this card still offers remarkable performance compared to current mid-range cards.
The new middle-class cards from NVIDIA and ATi also show us an interesting development: the new, highly praised features like DirectX 9, etc. have one major drawback - they require a lot of circuitry and, consequently, lots of transistors. Less expensive versions based on toned-down high-end chips will inevitably have to be pared down drastically to reduce the transistor count, thereby increasing chip yield and reducing manufacturing cost. Obviously, this will just as inevitably mean reduced performance. In the past, however, most chips of a product line differed only in clock speed, which was no great surprise, considering their moderate transistor count. A GF4 Ti chip only needed about 63 million transistors. If a card with 70-80 million transistors is supposed to incorporate DX9 features, the manufacturer is going to be forced to cut corners somewhere else. From the chipmakers' perspective, the prime candidates are the transistor-hungry DX9 texture and vertex shader units. The reduction comes at a different price though - fewer vertex and texture shaders also mean lower overall performance, especially in older games.