For quite a long period of time Silicon Integrated Systems Company, which is pretty well-known as a chipset manufacturer, has been producing graphics chips as well. Just like the mainboard chipsets from SiS, which have never been known as the fastest and feature-rich solutions, but attracted the manufacturers' attention due to their extremely low price, the graphics cards based on SiS chips also won the customers' hearts due to their low cost in the first place rather than to their performance. Of course, SiS is evolving, and every next product they develop is certainly better than the previous one. However, Intel, VIA and AMD in the chipset market and NVIDIA and the rest in the graphics market have always been at least one step ahead.

However, the situation has changed drastically lately. New SiS chipsets turned out really successful: SiS735 has already proven to be one of the best among all the currently available DDR chipsets for Athlon CPUs (see our Elitegroup K7S5A Review), SiS645 tends to be a success among the Pentium 4 chipsets…

In the graphics field SiS has also made some progress recently. While NVIDIA and ATI were trying to beat one another in the High-End and Mainstream market segments, SiS introduced a new low-cost SiS315 graphics chip for the Value market. To tell the truth, we believe that this chipset may turn out a very important milestone for SiS's career in the graphics chips market.

What made us think so? Well, to answer this question let's take a closer look at SiS315 and its features.

SiS315 Specifications

Here is a scheme of SiS315 architecture:

• 0.15micron manufacturing technology;
• 64bit/128bit graphics memory interface;
• Up to 128MB SDRAM/DDR SDRAM and SGRAM/DDR SGRAM;
• AGP 4x interface;
• 166MHz chip frequency;
• 166MHz memory frequency;
• 375MHz RAMDAC frequency;
• Maximum supported resolution: 2048x1536;
• Built-in Direct Draw and GDI 2000 accelerator;
• Built-in motion compensation logic for DVD playback;
• 2 pixel pipelines with one texturing unit on each;
• Bi-linear and tri-linear filtering support;
• Bump mapping support, namely: Dot3 and Emboss;
• Up to 2048x2048 texture size;
• S3TC/DXTC texture compression support;
• 2nd generation hardware T&L engine;
• HyperFSAA full-scene anti-aliasing.

Moreover, as you can see from the picture above, SiS301 companion-chip adds some multimedia features, such as TV-Out and stereoscopic glasses support, the possibility to connect digital displays via the DVI interface. There are the following combinations available:

• Analog display + analog display;
• Analog display + digital display (DVI);
• Analog display + TV (S-Video Out).

In case of two displays connected, there are two working modes available: desktop extension and cloning.

Judging by the listed specs, SiS315 doesn't yield too much to GeForce2 MX 200/400 from NVIDIA. Note that though the latter requires additional chips from other manufacturers to be integrated onto the card in order to implement the TV-Out and DVI support, SiS315 solution requires only SiS301 companion-chip developed specifically for these particular needs.

We will dwell on SiS315 features later on in the article, and now meet the graphics card built on it, which we managed to get hold of.

C.P. Technology Evil Sam Graphics Card on SiS315: Closer Look

Evil Sam graphics card from C.P Technology is shipped in a not very large radically black box:

The card itself is designed on a yellow PCB and features a DVI-connector for digital displays and an S-Video Out.

   

The card is built on SiS315 core and SiS301 companion-chip:

The core is equipped with a passive pin-heatsink, which seems to be quite enough to provide proper cooling, because we didn't discover any significant temperature increase during our tests.

The card is provided with 32MB SDRAM with 128bit bus. The memory chips are made by Samsung and feature 5ns access time:

Both, the graphics core and the memory work at 166MHz.

Testbed

Our testing system was configured in the following way:

• AMD Athlon 1.2GHz CPU (133MHz FSB);
• A7V133-C mainboard (VIA KT133A);
• 256MB NCP PC133 SDRAM;
• Fujitsu MPE3084AE 8.4GB HDD.

The software used included:

• Windows 98 SE build 4.10.2222 A;
• DirectX8.0.

We compared the performance of our SiS315 based graphics card with MSI StarForce 828 based on NVIDIA GeForce2 MX 200 and equipped with 32MB memory and SUMA Platinum K1 32MB on PowerVR/STM KYRO chip.

We used the following drivers:

• GeForce2 MX 200 based card - Detonator XP driver ver. 21.85;
• SiS315 based card - driver ver. 2.02.05;
• PowerVR/STM KYRO based card - driver ver. 8.162.

Overclocking

For overclocking experiments we resorted to PowerStrip utility ver. 3.0beta build 115. Despite the fact that this card features graphics memory with 5ns access time, which should work fine at 200MHz, the maximum frequencies for the board to run stably made only 190MHz for the chip and 190MHz for the memory. It's hard to explain what the reason for such low overclocking potential is: the card layout, the chip architecture peculiarities or just our bad luck (we could have got a badly overclocking sample card).

The performance of the overclocked SiS315 is included into the diagrams for Quake3 Arena in the "3D Performance" section of this review.

2D Performance and Image Quality

To test the performance of our graphics card in 2D we used Business Graphics and High-End Graphics from ZD WinBench99 benchmark set:

As you can see from the diagram, SiS315 performs in 2D applications almost as fast as GeForce2 MX 200 leaving KYRO far behind. Here is a more detailed illustration made in the High-End Graphics test, which proves our conclusion:

As for the 2D image quality provided by SiS315, we liked it very much. Only when we set the resolution higher than 1280x1024 we could notice some blurring effect on our ViewSonic P775 monitor.

Pipelines, Texture Filtering, HyperFSAA and Other

SiS315 features only one texturing unit per pipeline that is why the chip has to unite the pipelines to perform multitexturing. As a result the performance drops, because the chip draws not two pixels per clock but just one pixel. The same algorithm NVIDIA used in its TNT2 and TNT2 Ultra chips some time ago. Now you will hardly find a game without any multitexturing that is why the architecture of the kind will in no case make SiS315 a competitor to GeForce2 MX 400 and faster solutions.

But this is just the beginning. Following the good traditions set up by NVIDIA in its TNT/TNT2 and by S3 in its Savage2000, SiS315 doesn't know to perform tri-linear filtering together with multitexturing.

For example, in Quake3 Arena the graphics quality settings where we enabled tri-linear filtering, didn't work and we got real tri-linear filtering only having disabled multitexturing by typing ""/r_ext_multitexture 0" in the console. Moreover, the texturing (this is usually the base texture and the lightmap) is carried out in two passes, so that the performance gets twice as low, correspondingly. Frankly speaking, this situation is absolutely unacceptable even for a value graphics card.

In case the MIP-level coloring is on (""/r_colormiplevels 1" in the console) the scene in Quake3 Arena looks like this (bi-linear filtering on the left, tri-linear filtering on the right):

Bi-linear filtering		Tri-linear filtering

But this is again not everything we wanted to mention. Hanging around Quake3 levels looking for artifacts we discovered that bi-linear filtering was performed by SiS315 not in the proper way. Having come close with the nose nearly pressed to the wall, we saw some rough edges, which didn't belong to the actual picture:

Having reduced the viewing angle with "/cg_fov 5" command, we observed the wall attentively and noticed that there was some kind of approximation used instead of the bi-linear filtering, which looked much more like dithering:

To tell the truth, this effect is not that visible on the base textures of the objects and doesn't arouse any uncomfortable feeling if noticed. However, on lightmaps it inevitably catches your eye:

And if you get closer to the lit surface, it will be very clearly seen:

However, things are not that bad. SiS315 can perform proper bi-linear filtering as well. On the next screenshot you can see normal bi-linear filtering and the mentioned approximation together:

And now a close-up:

We failed to find out what makes SiS315 use approximation instead of bi-linear filtering. The only thing we managed to prove was that SiS315 always applied fully-fledged bi-liner filtering to transparent textures.

What's the use of this approximation? What are its advantages? Is it aimed at saving the chip's computing resources? In fact, bi-linear filtering doesn't require too much from the chip. Maybe then this approximation serves to reduce the number of samples requested? Maybe it's true, because the texturing unit needs 4 texturing samples to perform normal bi-linear filtering, and in case of SiS315 there seems to be only one sample available. Anyway, the texturing cache of SiS315 should reduce (almost to null) the difference between the number of memory requests in case bi-linear filtering is on and in case it is off. Therefore, the use of this approximation is not so justified.

There is only one supposition remaining: the texturing cache of SiS315 is very small and inefficiently organized, so that the developers had to resort to tricks like that when implementing bi-linear filtering...

Here we would like to stop the theoretical discussion about the texture filtering stuff, and suggest passing over to more pleasant things.

The press-releases about the announcement of SiS315 mentioned FSAA (full-scene anti-aliasing) support called HyperFSAA. Unfortunately, the driver version 2.02.05, which we used for our tests did not allow enabling FSAA, though in the Windows Registry we could find some mention of FSAA:

After a number of manipulations with the FSAA, FSAAMode in Direct3D and GL_FSAA, GL_FSAAMode in OpenGL, we managed to enable FSAA on SiS315 and the results appeared beyond our bravest expectations. FSAA on SiS315 turned out not even a bit worse than MSAA (Multi-Sampling anti-aliasing) of GeForce3. And this feature belonged to the card, which couldn't be even compared to GeForce3 as far as price and other characteristics go!

However, let's start from the very beginning. We tested the FSAA in Homeworld: Cataclysm game, which looks just perfect with FSAA on and can be run in OpenGL as well as in Direct3D without any problems. For each graphics card tested we took a screenshot from the game paying attention to two things: how well the details of our Hive Frigate are smoothed and how the stats is displayed in the lower right corner of the screen. Click on the first picture to enlarge (the other two are fragments of the image already enlarged as 3:1 for your convenience). So, first comes the scene without FSAA:

No FSAA: SiS315

When we set FSAA=1, FSAAMode=1 in Direct3D, SiS315 implemented anti-aliasing, which looked more like MSAA 2x by NVIDIA GeForce3.

In order to let you "feel the difference" we will try to compare the FSAA implementation on SiS315 with FSAA 2x, which looks absolutely the same for all NVIDIA graphics cards based on GeForce2 and GeForce2 MX. On the left you see the results obtained on SiS315, and on the right - on GeForce2 MX 200:

FSAA 2x: SiS315		FSAA 2x: GeForce2 MX 200

Looks impressive, eh? Of all the today's graphics accelerators only GeForce3 can perform anti-aliasing "only where necessary", i.e. only on the polygon edges. What mechanism works in SiS315 is still a mystery, but the outcome is great.

However, our tests show (and you will see it later on) that the performance of SiS315 with FSAA on (with HyperFSAA 2x, as we will call it) is very close to that of GeForce2 MX 200 with the regular FSAA 2x. So, it makes us believe that HyperFSAA of SiS315 is still based on the common supersampling and not on the multisampling like in NVIDIA GeForce3.

We didn't manage to enable FSAA 4x in Direct3D, but in OpenGL when we set GL_FSAA=1, GL_FSAAMode=1 we succeeded.

To compare the results obtained with the FSAA 4x by NVIDIA we would like to offer you a screenshot taken on GeForce2 Ti (GeForce2 MX 200 lacked graphics memory to enable FSAA 4x, unfortunately):

FSAA 4x: SiS315		FSAA 4x: GeForce2 Ti

As you can see, the situation repeats: SiS315 effects anti-aliasing only in those places where it is required, that is on the polygon edges, leaving the text unchanged. Since SiS' engineers didn't offer a name for this anti-aliasing method we dare call it HyperFSAA 4x.

GeForce2 Ti didn't save the text from FSAA 4x and it appeared completely washed-out just like the entire scene.

And finally comes the dessert :-) HyperFSAA from SiS turned out cool enough to compete with Quincunx from NVIDIA!

Having set GL_FSAA=1, GL_FSAAMode=0 in OpenGL we enabled a new type of anti-aliasing, which effect can be compared with that provided by Quincunx from NVIDIA. On the left you can see a new HyperFSAA variation on SiS315 and on the right - Quincunx on NVIDIA GeForce3:

FSAA 2x + Filter: SiS315		MSAA Quincunx: GeForce3

Just like Quincunx on GeForce3, this method works as Blur filter in Photoshop. The image obtained on SiS315 looks even more blurred especially on the polygon edges. However, despite the similarity in the effect made, the algorithms applied by these two graphics cards are totally different.

Quincunx from NVIDIA is based on a tricky method of sampling from the enlarged frame buffer, which provides this "blurred" effect, because the color of the pixel is obtained using the samples corresponding to the neighboring pixels.

SiS315 simply smoothes the image obtained with HyperFSAA 2x in the frame buffer, just like the Blur filter from Photoshop.

You will see the proof of our point if you look attentively at the part of screenshot with the text: all letters got completely blurred as well as the polygon edges of our Hive Frigate. The typical edges and the fact that they turned out more washed out than the text, as well as the performance drop caused by this anti-aliasing method also prove that we deal with HyperFSAA 2x with additional filtering in the frame buffer. Let's call it HyperFSAA 2x + Filter. It seems that the filter is selected so that to provide the same image on the polygon edges as by GeForce2 FSAA 2x (compare the pics and you'll see what we mean). However, the polygon edges on SiS315 get much more blurred than on GeForce2 in this case.

We think that it is too much blurring for Homeworld: Cataclysm game. HyperFSAA 4x is more than enough here. However, it will be just the right thing for some dynamic games, such as race simulators, where blurred images add more realism to the scenes. Let's only hope that the performance of SiS315 will be high enough for good gaming in this case.

However, unfortunately, FSAA implementation by SiS315 is not absolutely impeccable. In some cases HyperFSAA 2x + Filter doesn't work properly: for instance, if the frame buffer doesn't get refreshed, the image stored in it is blurred more and more. A typical example is the menu page in Homeworld: Cataclysm game. When the image gets into the frame buffer the HyperFSAA 2x + Filter processes it once. Then SiS315 keeps processing it anew for every ongoing frame, because the frame buffer is not refreshed. As a result, the menu starts melting right in front of your eyes, getting more and more blurred. This phenomenon illustrates just perfectly what SiS315 does to the frame buffer applying HyperFSAA 2x + Filter method. Here are two screenshots of the mentioned menu. The right one was taken 30 seconds after the first one:

  

It is very likely that SiS developers decided not to allow enabling HyperFSAA in the meanwhile, because of these particular problems. Once the problem is eliminated, HyperFSAA will become available for everyone.

And in conclusion to our investigation we would like to comment on the following. Many people think that if the graphics card costs some trifling sum of money and cannot compete with the latest High-End graphics cards in terms of performance, then it doesn't need any specific features, such as FSAA, typical of powerful accelerators. However, we would like to totally disagree with that, and insist that any graphics card can prove an excellent product if the proper application field is found.

For example, in our case the specific feature is HyperFSAA. Of course, SiS315 will prove quite slow in Quake3 Arena, for instance, with the enabled HyperFSAA 4x. However, you will be able to enjoy all the beauty of such "calm" games as Homeworld: Cataclysm even with HyperFSAA 4x.

All in all, you shouldn't disregard the features offered by the product developers, especially if they are so successfully implemented as FSAA on SiS315.

3D Performance

And now it's high time we passed over to the tests. Bearing in mind that the performance of SiS315 drops down significantly once tri-linear filtering is enabled, the tests in Quake3 Arena were run with the maximum texture quality settings and bi-linear filtering. Texture compression was enabled:

As you can see from the diagram, the fastest one here appeared KYRO. No wonder, as it feels quite at home in tests like that without tri-linear filtering.

Despite lower fillrate of SiS315 compared with GeForce2 MX 200, our hero managed to outperform the latter in 16bit color and in 32bit color modes. It is probably the cut down 64bit memory bus of the MX 200 solution that tells even in 16bit mode.

Now we would like to check the graphics cards performance in Quake3 Arena with FSAA on:

Well, almost 50% of the performance is sacrificed for the sake of HyperFSAA 2x + Filter method, which proves our supposition that it is based on HyperFSAA 2x. The additional smoothing of the frame buffer hardly tells on the performance.

The performance drop with HyperFSAA 4x is much higher and is close to the performance drop percentage shown by the other testing participants working with the enabled FSAA 4x.

Although SiS315 yields to KYRO here, it manages to surpass GeForce2 MX 200. At the same time, the FSAA quality provided by SiS315 is undoubtedly beyond any competition.

Taking into account that these cards are positioned as Value solutions, we tested them only in Low Detail mode in 3DMark2001.

KYRO wins all the laurels here, which is quite logical: when we compared the performance of KYRO and KYRO2 with that shown by GeForce2 GTS, we mentioned that despite weaker KYRO architecture these solutions sometimes manage to leave GeForce2 GTS behind. No wonder that KYRO easily defeated such rivals as SiS315 and GeForce2 MX 200.

As we have expected, SiS315 fell behind GeForce2 MX 200 in 16bit modes because of twice as few texturing units. In 32bit modes it managed to beat the competitor due to twice as fast graphics memory bus.

The last 3D benchmarks we would like to focus on are synthetic tests from 3DMark2001 set:

Well, what can we say here? Yes, SiS315 has a T&L unit and it works faster than the software emulation by Athlon 1.2GHz. However, its performance is still much lower than that of the corresponding unit by GeForce2 MX 200.

Judging by the fact that SiS315 failed to win even with the software emulation of the T&L unit and vertex shaders, it suffers some problems with the drivers: the polygons are probably transferred to the core at a much slower speed than by GeForce2 MX.

The fillrate test didn't reveal anything principally new: without the multitexturing, all the three testing participants - SiS315, GeForce2 MX 200 and KYRO - appear in equal conditions, as all of them have 2 pipelines. In this case, it is the working frequencies and the efficiency of memory operation that matter. SiS315 working at higher core and memory frequencies than KYRO (166MHz/166MHz against 115MHz/115MHz) turned out ahead.

As soon as we enable multitexturing, MX 200 appears an indisputable leader in 16mode, as it has 4 texturing units and even its narrower 64bit memory bus doesn't spoil the whole picture.

But in 32bit color the memory bus can't help telling negatively on the performance and the victory goes over to KYRO and SiS315.

Practical Gaming with SiS315

Remembering that SiS315 is an offspring of the company, which is famous for constant problems with the drivers, we couldn't help taking a look at this card in real applications: in games. The results appeared surprisingly good: we faced nearly no problems when playing. Here are the games we took a look at:

American McGee's Alice

No comments here. The game is based on the modified Quake3 engine, so we didn't expect any problems to arise.

Earth 2150

This is a strategy working in Direct3D. It looks quite OK on SiS315. The only thing we could point out here is the noticeable boundaries between the MIP-levels, which turn out so visible because of no tri-linear filtering.

Giants: Citizen Kabuto

This 3D game makes active use of T&L, Dot3 bump mapping and seems to be alright at first sight. However, if you click the picture to enlarge you will notice that the shadow dropped by the flying beast is square.

Homeworld: Cataclysm

How could we leave out Homeworld and Homeworld: Cataclysm games, which we used to be so addicted to some time ago! :-) In Direct3D everything went on well, however, in OpenGL we discovered some problems: the stars were covering the spacecrafts:

Although this problem is typical of many graphics accelerators working in OpenGL (not only SiS315 suffers from it), we still can't omit this drawback.

Independence War II: Edge of Chaos

SiS315 proved quite worthy "at the edge of chaos". It is only its notorious bi-linear approximation that catches your eye, and the text in the menu:

Look, in some lines the letters are on different levels.

Return to Castle Wolfstein

No remarks to be made.

Soldier of Fortune

We don't have any significant comments here, actually. Only sometimes bi-linear approximation on SiS315 looks as if absolutely no texture filtering had been made:

Conclusion

Well, SiS315, the latest graphics solution from SiS, is undoubtedly a very diverse product.

On the one hand, a new graphics chip is a significant move forward for SiS Company: the drivers have become much better, there appeared HyperFSAA anti-aliasing, the 2D image quality is no longer scary.

On the other hand, we have this notorious bi-linear approximation, some bugs in the drivers…

Nevertheless, summing up we could say that SiS315 is a good combination of low cost, pretty good performance and some interesting features. And the discovered drawbacks will not prevent it from becoming a worthy competitor to KYRO and GeForce2 MX 200 in the value graphics market.

Highs:

• Dual-monitor configurations support;
• DVI-, TV-Out;
• T&L Unit;
• High-quality FSAA;
• Good gaming performance;
• Low cost.

Lows:

• Bi-linear approximation;
• No simultaneous support for tri-linear filtering and multitexturing;
• Low overclockability;
• Raw drivers.