by Alexey Stepin
04/19/2005 | 07:19 PM
You may have read our recent preview of the new graphics processor from S3 called GammaChrome (see our article called S3 GammaChromeS18 Graphics Processor Preview). Today we’re going to examine this chip and the graphics card based on it in more detail.
The documentation we received from S3 Graphics says there will be two versions of the product – GammaChrome S18 Pro and GammaChrome S18 CE – differing in the clock rates and the amount of graphics memory on board. The senior, S18 Pro model is expected to carry 256MB of memory and work at 400/300 (600DDR) MHz frequencies; the junior, S18 CE model comes with 128MB of memory and works at 300/300 (600DDR) MHz. The GammaChrome S18 Ultra version that we have previously mentioned on our site is missing in this list.
We received an engineering sample of the GammaChrome S18 Pro for our tests. Packed in an antistatic bag, the card was only accompanied with two discs with drivers from S3. The device didn’t even have the mounting bracket you fasten a graphics card in your system case with, but it did have two paper tags, one tag showing the frequencies of the sample as 350/350MHz. The tag proved to be wrong eventually as the real frequencies of the GPU and memory were 400MHz and 350 (700DDR) MHz, respectively.
The official documentation from S3 Graphics claims the final version of the GammaChrome S18 Pro is going to be clocked at 400/300 (600DDR) MHz, so the GPU of our sample worked exactly at the declared frequency, while the memory was clocked slightly higher than specified.
Now let’s have a closer view of the card.
The GammaChrome appears to be a complex device, with numerous small elements scattered all around its deep-blue PCB. Like the GeForce 6600 GT, the GammaChrome has a 128-bit memory bus, but the former works at much higher clock rates. So, why all this complexity? The S3 engineering team may have encountered instability in the operation of the graphics card and had to make the wiring of the PCB more complex. If it’s true, this fact also explains why the company has so far refused to launch the GammaChrome S18 Ultra whose core was going to work at 500MHz.
The power circuitry is, however, simple, since the GammaChrome isn’t a very hungry consumer of power. The card doesn’t carry additional power connectors. The back side of the PCB has even more small components than the front side: there are four more memory chips here with all their accompaniments as well as the load-bearing elements of the power circuit. Besides that, you can see the S3 Graphics and the PCI Express logos here.
The graphics card offers the usual selection of connectors: D-Sub, S-Video and DVI-I, the latter provided by a TMDS transmitter from VIA (the VT1632A chip, which supports resolutions up to UXGA or 1600x1200). The S-Video output is actually a universal connector that display devices with S-Video, RCA or YPbPr inputs can be attached to, via a special splitter.
Our sample of the S3 GammaChrome was ventilated with a modest aluminum cooler with a small fan and a sticker with the company’s logo. Very thick white-color thermal paste was the thermal interface. The fan doesn’t have a tachometer and doesn’t seem to be controlled depending on the temperature. This cooler seems to be appropriate, considering the low power consumption of S3’s graphics processors, but the final version of the GammaChrome is going to bear a bigger cooler. The small cooler on our sample was fastened with two standard plastic spring-loaded clips and concealed the heart of the device – the GammaChrome processor.
As you see, S3 followed the example of ATI and NVIDIA and switched to the open-die type of packaging. The size of the graphics core is small as it contains only four pipelines. The marks on the chip say this sample of the GammaChrome was manufactured the second week of 2005, i.e. in the middle of January. There’s a traditional protective frame around the chip to prevent the cooler from misaligning and the core from chipping.
The graphics card bears eight 256Mbit GDDR SDRAM chips from Samsung with an access time of 2.86 nanoseconds. This means the chips are rated for 350 (700DDR) MHz frequency and that’s exactly the frequency they are clocked at on this card. The chips remain rather cool at work, so we don’t consider the lack of heatsinks on them as a drawback.
We got two CDs from S3 with the version beta-1904-15.14.10e driver. After the installation, there appeared the standard set of S3 tabs in the graphics adapter’s properties window. We described them all in our DeltaChrome S8 review (see our article called The Return of S3: DeltaChrome Graphics Card Review) and the S3Config OGL tab in the DeltaChrome S8 Nitro review (see our article called S3 DeltaChrome S8 Gets Nitro Acceleration: Review of the Revamped S8). As for new options, you can now enable 4x full-screen anti-aliasing (the DeltaChrome was limited by FSAA 2x), but the GammaChrome still uses the resource-consuming super-sampling method that leads to a tremendous performance hit in games.
There were no errors during the installation of the drivers, but we did have problems after that. We don’t know what exactly to blame – the imperfect drivers or some defects in the design of the graphics card itself – but errors and hang-ups occurred in many games. For example, the popular shooter Half-Life 2 would report the following:
Another popular game, Far Cry, also had image artifacts or invoked the notorious “Blue Screen of Death”. Games like Splinter Cell: Pandora Tomorrow and The Chronicles of Riddick wouldn’t start up at all; Aquamark3 would hang the computer up, too. Similar crashes and hang-ups occurred in a number of other games, with the following error message:
We can’t call this stable operation. We still hope, however, that these defects will have been corrected and the drivers improved by the time graphics cards on S3’s new GPU come to retail.
According to S3 sources, these games run fine on their systems, so we will work closely with S3 guys to find out what the problem was on our side. Stay tuned for an update.
We didn’t overclock the card: neither RivaTuner nor the version of PowerStrip we had at hand was compatible with the GammaChrome. The device was quite expectedly quiet – its noise was simply inaudible against the clamor from the rest of the system components. But keep it in mind that the final version of the GammaChrome is going to be equipped with a different and maybe louder cooler.
The image quality in 2D applications was up to the mark in all the resolutions despite the card’s status of an engineering sample. But unlike other graphics cards, the GammaChrome S18 wouldn’t permit to set a resolution above 1600x1200@85Hz, although our laboratory monitor supports display modes up to 2048x1536@60Hz. That’s a curious fact, but a fact nonetheless. At least we have this situation with the current drivers from S3 Graphics.
Like we always do when encounter a new graphics processor, we performed a series of theoretical tests to reveal the potential of the S3 GammaChrome in comparison with same-class solutions from other manufacturers. We’ll present the results of these tests below, after the description of our testbed.
We ran our tests on the following testbed:
We tried the newcomer at playing files of different formats (WMV HD, DivX and DVD). These tests were run on an i925X-based platform with a 3.60GHz Intel Pentium 560 processor.
For the HDTV test we took the Magic of Flight clip encoded in 1080p WMV HD format with a data rate of 8 megabits per second. You can download it here. So, the RADEON X600 PRO is the slowest in this test, while the S3 GammaChrome S18 is just a little behind the GeForce 6600 which sports a second-generation PureVideo video-processor.
To check out the CPU load during DivX playback we used a file encoded at 640x480 resolution. We got much more curious results here: the S3 graphics card had the highest minimal CPU load level when the file was being reproduced through the standard Windows Media Player, but the maximum peak of the load was the same as with the GeForce 6600. When we used the InterVideo WinDVD version specially optimized for the GammaChrome’s Chromotion engine, the min CPU load level went halfway down, while the maximum level remained the same. The RADEON X600 PRO has a smallest difference between the max and min CPU load levels, and it wins this test if we compare the results by the max CPU load.
We took a Starship Troopers DVD for the next test. The S3 GammaChrome wins this check when playing the movie through Windows Media Player. For some unknown reason the CPU load is considerably higher with the WinDVD player than in the previous test. The GeForce 6600 occupies the second place, followed by the RADEON X600 PRO.
So, the new chip from S3 displays its exceptional talent for video playback and makes a worthy rival to the NVIDIA GeForce 6600 GPU which is equipped with a special video-processor specially intended to reduce the CPU load.
The widely known Marko Dolenc’s Fillrate Tester helped us measure the scene fill rate parameter.
The measurements of the fill rate with enabled color and Z-value writes hint at high efficiency of the GammaChrome GPU, particularly its caches and memory subsystem. The GammaChrome’s fill rate diminishes almost like the RADEON X600 PRO’s as there are more textures to be mapped, but the S3 card is anyway a little better. This outcome may be due to more efficient pixel pipelines or to the higher memory frequency of the latter.
It’s similar to what we have seen above: the S3 GammaChrome S18 is more efficient than the RADEON X600 PRO, their core frequencies being the same but the memory frequency on the S3 card being slightly higher.
Relying on the NV4x architecture that allows writing a double number of Z values per clock, the GeForce 6600 leaves the rest of the test participants far behind. But the GammaChrome, though it doesn’t have UltraShadow II technology, is quite efficient with its Z buffer, too. It is more than by half faster than the RADEON X600 PRO.
Marco Dolenc’s Fillrate Tester can also check out the speed of the pixel processors. Below are the results we got with enabled color and Z writes:
The performance of the GammaChrome with simple pixel shaders is quite expectedly similar to that of the RADEON X600 PRO. It’s worse with the more complex per-pixel-based lighting where the new GPU is even slower than the RADEON X300.
Take note that the GeForce 6600 works faster in the half-precision mode, but its performance is anyway much better than of the rest of the participants of the test even at 32-bit precision.
When Z writes are disabled, we again see the GammaChrome S18 being slower with version 2.0 pixel shaders than the RADEON X300 with its 325MHz core frequency and the same number of pixel pipelines.
The picture is different in the contrary case – with disabled color writes and enabled Z writes. The diagram is indicative of high efficiency of the GammaChrome processor as it works with the Z buffer, especially when executing simple shaders.
Alas, the performance of the GammaChrome with pixel shaders isn’t as high as we might wish. The GammaChrome is no match to ATI’s solutions, not to mention the GeForce 6600 with its twice as many pixel pipelines, in tests which emphasize the pixel shader speed. The S3 card wins only three tests against the RADEON X600 PRO and six tests against the RADEON X300. The GammaChrome is especially weak at processing math1ematics-heavy pixel shaders like Dot Product Bump Mapping + Specular + Reflection or Wood.
The GammaChrome isn’t speedy even with simple version 1.1 pixel shaders – the RADEON X300 is faster in this test.
It’s somewhat better with version 1.4 shaders – the GammaChrome S18 is ahead of the RADEON X300 and closely following the RADEON X600 PRO.
The new product from S3 is sluggish with version 2.0 pixel shaders, too. It is just a little faster than the RADEON X300 here.
3DMark05 supports all existing versions of pixel and vertex shaders, including 3.0. Despite S3’s claims about support of extended pixel and vertex shaders, the benchmark found the GammaChrome S18 capable of executing version 2.0 shaders at most. The RADEON X600 PRO and RADEON X300 worked in this mode, too, while the GeForce 6600 used version 3.0 shaders. The pixel shader speed test from the 3DMark05 suite is much more complex than the analogous test from 3DMark03. Part of the cliff from the third game test is being rendered here, and this puts a heavy math1ematical as well as textural burden on the GPU.
As we already know from the results of Xbitmark, the GammaChrome S18 isn’t much versed in math1ematics. So, the twofold advantage of the RADEON X300 isn’t surprising at all. That said, the new graphics processor from S3 may have some issues with games considering the ever-growing complexity of pixel-shader-based special effects we see in modern gaming applications.
The performance of the GammaChrome’s four vertex processors roughly matches that of the two vertex processors of the RADEON X600 PRO, but the performance of the S3 card falls almost to the level of the RADEON X300 as soon as 1024x768 resolution. This test is rather out-dated, though. Let’s see what we have in more up-to-date ones.
The GammaChrome S18 benefits by its four vertex processors in the analogous test from 3DMark03 – this graphics card can’t catch up with the GeForce 6600 but leaves both cards from ATI Technologies behind.
The Simple Vertex Shader test from 3DMark05 isn’t as simple as its name implies. Here, transformation and lighting are performed over six models each of which consists of over a million vertexes.
The S3 GammaChrome S18 and the NVIDIA GeForce 6600 GT have four and three vertex processors, respectively. The performance of the two is similar in this test, though. This is a confirmation of our point that the efficiency of a vertex processor from S3 is lower than the efficiency of a vertex processor from NVIDIA, the clock rate being the same. Note also that the graph of the GeForce 6600 is almost a straight line, while the graph of the S3 GammaChrome S18 Pro shows a bigger dependence on the resolution.
The second vertex-shader-related test from the 3DMark05 suite is much more difficult than the first one: grass is being transformed and illuminated here. Each blade of grass is rendered independently, so the vertex processors of the graphics card experience a tremendous stress. The four vertex processors of the GammaChrome S18 show their best in this test: the card outperforms the RADEON X600 PRO and the GeForce 6600 in low resolutions. But in 1600x1200 the performance of the S3 card sinks to the level of the RADEON X600 PRO which has only two vertex processors.
Xbitmark can check out the speed of the vertex processors, too, so we ran this test to give another try to the GammaChrome.
The test, however, refused to work correctly with this graphics card and yielded extremely low results in all the modes. There seem to be some early driver issues here, but according to S3 sources, there are updated drivers that may remedy the problem.
Several tests from the 3DMark2001 suite can be employed to test the geometry processing speed of the graphics card, particularly the tests with one and eight light sources as well as the point sprites test.
The GammaChrome S18 performs like the RADEON X600 PRO in the single-light-source test, lagging somewhat farther behind in higher resolutions.
The card from S3 is ahead of the RADEON X600 PRO in the test with eight light sources – the GammaChrome’s four vertex processors contribute to this performance. The absolute winner is the GeForce 6600 since its core contains special units for an efficient execution of older T&L-using applications.
The Point Sprites test doesn’t only load the vertex processors, but also requires the graphics card to have a high texturing speed and a fast memory subsystem. These units of the GammaChrome S18 work fine, but the drivers from S3 are early versions. So, the new card delivers the performance of the RADEON X300.
The previous, DeltaChrome S8 processor was bad at rendering relief with the EMBM method, and the new GammaChrome hasn’t improved since then. It can only compete with the RADEON X300 which works at lower frequencies.
The GammaChrome S18 did better with another relief-rendering method, Dot3, but couldn’t reach the level of the RADEON X600 PRO. The GeForce 6600 is unrivalled here – as we said above, NVIDIA’s solution works most efficiently when graphics cards without special support of fixed T&L falter.
The Ragtroll test shows the balance of the “CPU-driver-GPU” chain since the CPU is calculating the physics of the scene, while the GPU processes the geometry. The S3 GammaChrome is better than the RADEON X600 PRO here, especially in high resolutions. The reason is simple: the GammaChrome has four vertex processors against the opponent’s two and has a higher memory frequency.
The results of the theoretical tests suggest that the GammaChrome S18 is capable of rivaling such solutions as RADEON X300 or X600 PRO, but the latter are going to be faster in modern games since the S3 card is rather slow at executing complex pixel shaders. The GammaChrome S18 is better at processing geometrical data, so the decision of S3 Graphics to give four vertex processors to their new GPU seems right to us.
Our final judgment will come after testing the card in real games, and that’s what we’re going to do right now.
We have already described the configuration of our testbed, so here’s a list of graphics cards that took part in our tests along with the card from S3:
We set the “Detail level adjust” option to 0 (which means the highest quality mode) in the S3 driver for both DirectX and OpenGL. Other Direct3D options were left intact, that is “Auto”. We enabled the optimizations in the NVIDIA ForceWare driver, save for the “Anisotropic mip filter” optimization. The “CATALYST A.I.” option was set to “Standard” in the ATI Catalyst driver, and the “Mipmap Detail Level” to “Quality”; we used the “Quality” mode. The “VSync” option was disabled in all the drivers.
If the game permitted to enable full-screen anti-aliasing and anisotropic filtering, we used this option. Otherwise we forced the necessary mode from the driver. The tested solutions belong to the value-mainstream class, so we decided not to test resolutions above 1024x768 in the “eye candy” mode. We remind you that the S3 GammaChrome S18 uses super-sampling as the way of doing full-screen anti-aliasing, so we shouldn’t expect it to perform fast in the “eye candy” mode.
We selected the highest graphics quality settings in each game. The following games were used:
First Person 3D Shooters:
Third Person 3D Shooters:
This list doesn’t include those games and benchmarks which were unstable or prone to hang up on the GammaChrome graphics card, namely:
The GammaChrome takes a nice start, leaving the RADEON X600 PRO behind in all the resolutions. The efficient use of the Z buffer and the availability of four vertex processors must be the main factors contributing to the success. But on the other hand, the GammaChrome yields just a paltry 21 frames per second in the lowest resolution – we can’t call this fast. The GeForce 6600 is the only card in this test session to run Doom 3 at an acceptable speed.
We didn’t use the “eye candy” mode with enabled anti-aliasing and anisotropic filtering because Doom 3 is an OpenGL-based application and the current driver from S3 doesn’t support full-screen anti-aliasing for this API.
There are no monsters on the d3dm4 map, so the frame rates are overall higher. There’s a wider gap between the S3 GammaChrome and the RADEON X600 PRO than in the previous case, but comfortable play is only ensured by the GeForce 6600.
The GammaChrome is slower than the RADEON X600 PRO and GeForce 6600 but faster than the RADEON X300 that works at lower frequencies.
As we had expected, the GammaChrome suffered a performance hit in the “eye candy” mode. It was twice as slow as even the RADEON X300. So, S3’s full-screen anti-aliasing has not that much of a practical value here: it may not work comfortably with this feature turned on.
The GammaChrome outperforms the RADEON X300 and provides enough speed in low resolutions on the Metallurgy level.
Again, the performance of the graphics card from S3 plummets down in the “eye candy” mode, which is only really playable on the GeForce 6600.
The low pixel shader performance of the GammaChrome is apparent here as the card even falls behind the RADEON X300, despite having a higher core frequency.
The newcomer is as fast as the RADEON X300 in Il-2, and even a little faster in high resolutions. Anyway, the efficiency of the GammaChrome might be better – our sample is clocked at 400/350 (700DDR) MHz, while the ATI card is clocked at the humble 325/200 (400DDR) MHz. We don’t test the FSAA+AF mode because the current OpenGL driver from S3 doesn’t support full-screen anti-aliasing.
This game uses pixel shaders extensively, but the GammaChrome isn’t the best executor of such shaders. It can only compete with the RADEON X300 here.
Multi-sampling is victorious over super-sampling in the “eye candy” mode: the GammaChrome has a very humble result, even in comparison with the RADEON X300.
The four vertex processors helped the GammaChrome to outperform the RADEON X300 in this game, but the deficiencies of the new product from S3 don’t allow it to reach the level of the RADEON X600 PRO, not to mention the GeForce 6600.
The S3 card falls behind the others once again in the “eye candy” mode. This time, however, it is close to the RADEON X300 which can’t be better in this test because of its slow memory.
Perimeter is a very demanding game and none of the tested cards ensured a comfortable frame rate even in 1024x768. As for the GammaChrome S18, it has the worst result among all the participating cards.
The GammaChrome S18 is slightly behind the RADEON X300 in this semi-synthetic test. The GeForce 6600 has an even smaller result due to the problems with NVIDIA’s cards on the PCI Express platform – this result shouldn’t be considered correct, of course.
Surprisingly enough, the GammaChrome S18 has got an excellent 3DMark03 score, being more than 500 points better than the RADEON X600 PRO. Where does this high total score come from?
The four vertex processors must be the main contributing factor to the advantage the GammaChrome has over the RADEON X600 PRO in the first test. The GeForce 6600 is on top, but don’t forget that it is an eight-pipelined solution, while the other cards have only four pixel pipelines each.
The S3 GammaChrome S18 is the slowest card in the “eye candy” mode due to the same reason – its way of doing full-screen anti-aliasing is too resource-consuming.
The second test usually runs faster on NVIDIA’s cards, but the GammaChrome isn’t the worst performer here, approaching the GeForce 6600 closely in 1600x1200. The new GPU from S3 is aided greatly by its four vertex processors as well as by its efficient use of the Z buffer. Besides that, it is equipped with memory clocked at 350 (700DDR) MHz – the highest memory frequency among the participating cards.
The advantages of the GammaChrome that helped it in the second game test outweigh the disadvantages of super-sampling, and the card gets the second place, after the GeForce 6600.
The third test resembles the second one, so the above-described situation is repeated here.
Like in the previous case, the GammaChrome remains superior over the RADEON X600 PRO even with enabled full-screen anti-aliasing.
Despite its low pixel shader performance, the GammaChrome equals the RADEON X600 PRO in the fourth test.
The new graphics card cannot stand its ground in the “eye candy” mode: it is slower even than the RADEON X300.
But the scores are correct – the GammaChrome is better in three tests and performs as fast as the RADEON X600 PRO in the fourth test, so it deservedly has a higher total score.
The GammaChrome S18 looks well in 3DMark03, too. Its result is low as it is, but higher than the RADEON X600 PRO has got.
The GammaChrome goes abreast to the RADEON X600 PRO in the first game test which depicts a typical scene from a first-person shooter.
The GammaChrome doesn’t have the same disastrous performance hit here as it had in many games as soon as we enabled full-screen anti-aliasing and anisotropic filtering.
The second test looks like a role-playing game, and the new graphics processor from S3 succeeds here: it is a little behind the GeForce 6600 and far ahead of both RADEONs.
The same goes for the “eye candy” mode.
The GammaChrome S18 is quite successfully crunching through the third game test with its sophisticated pixel shaders. It is still ahead of the RADEON X600 PRO, although not as far ahead as in the second test.
The GammaChrome is no worse than the card from ATI Technologies even with enabled full-screen anti-aliasing and anisotropic filtering. So, the total score of the S3 card seems well deserved.
The performance of our GammaChrome sample turned to be higher than that of the RADEON X300, but lower than that of the RADEON X600 PRO. In some cases the new card was faster than the X600 PRO, though. The value solution from S3 performed well in the video tests, ensuring a low CPU load even during playback of files in the WMV HD format. So, the GammaChrome S18 Pro has a high potential, and that’s nice.
The graphics card on the new GammaChrome S18 processor from S3 Graphics didn’t perform well in our today’s tests probably because of the early driver version and, probably, because of a few issues with the card design. As a result, the device was sometimes unstable – some games would freeze the system soon after their launch. We also saw the display mode reset to 640x480@4bit in 2D applications a few times.
As you may see from our benchmark results, the GammaChrome solution appeared relatively unstable in some tests. A graphics card turned out incompatible with a number of games, so it will be difficult for it to win the hearts of many users this way, no matter how technologically advanced it may be. I am pretty sure that the graphics cards manufacturers would rather prefer to wait until S3 Graphics finalizes this product before the GammaChrome comes to shops. It is of course possible that the problems are only peculiar to our sample of the card, and that the mass production solutions will run better and faster, but we can’t yet reject our suspicions about the early drivers.