SiS aims higher

While the market share of 6326 was a big achievement, the cheap tier of the performance segment was not. User expectations were increasing quickly as well. SiS prepared a response in the spring of 1999 and it was named 300. Still no fancy names, just "300". But just because the name is boring does not mean the chip isn't interesting. If the previous 6326 is taken as a baseline, the 300 represents the kind of generational leap that legends are made of. Yet, chances are you haven't heard about SiS300, not in any detail at least. In fact, this is the only graphics chip from 1999 I cannot find a single review of. And the cards are rare as well. The main question is therefore simply this: why so obscure?

3d engine

The SiS 300 is defined by the latest greatest 250 nm manufacturing process available. The clock reached 125 MHz and perhaps more importantly, it is a robustly single-cycle architecture. The fillrates are indeed close to 125 million pixels per second with many features enabled. But let's start at the beginning. The chip supports AGP 1.0 with side band addressing, not more (though it is ready for 1.5V). That is rather unfortunate timing since competition will begin to offer AGP 2.0 chips at the same time. But this performance class does not need a faster bus. At last, 32 pipeline stages are supported. Turbo Queue, the command FIFO in local memory, remains with sizes up to 512 kB. The setup engine is capable of triangle setup within 60 clocks, just like the 6326. If the speed is the same for all triangle parameters, the SiS 300 can enjoy higher triangle throughput only thanks to higher clocks. Then the rasterizer can generate shaded pixels of triangles and passes them to the texturing stage. Rectangular textures are supported in usual color formats. 4 and 8-bit texture palette ability remained for compatibility with older Direct3D titles. So far not much difference, only a slight evolution from the 6326. The texture cache however is now 8 kB large, sized for full speed even when trilinear filtering is used. And texture resolution up to 2k is also supported. I am happy to report the bilinear filter simplification of 6326 is gone. Single pass (but not single cycle) multitexturing is possible including emboss bump-mapping. Once texturing is done, the final stage of the engine applies fogging, blending dithering, and outputs given pixel to memory.

All the Z-buffer and scissor tests remained as well. The precision of the Z-buffer can be raised to 32 bits and framebuffer color depth of course also supports 32 bits. To do that, with trilinear texture filter, alpha blending and all the work traditional accelerator does, within 90 % of the theoretical maximum fillrates, the question arises: how do you satisfy the memory bandwidth requirements? And SiS simply answers: 128-bit bus, inside and out, double of 6326. Overall, this is a modernization of the 6326 that makes it wider, built for 32-bit rendering with trilinear texture filtering. On top of that rather advanced dynamic power-saving techniques are implemented and the result is a chip that can satisfy contemporary needs while being happy with an anemic heatsink. Is there a feature I would personally miss a lot? Yes, per-pixel mipmapping.

The card

HIS Graphic Force - a complete package for dual monitor/TV and stereoscopic 3D glasses.

One of the rare occasions of having an original box. Never mind the stated 166 MHz DRAM clock. It was a maximum according to the datasheet, up from a preliminary 133 MHz, so the memory controller ought to turn out fine. However, SDRAM memory chips are rated for 7 ns, making me fairly certain they are clocked at no more than 143.3 MHz. Most importantly the control panel says the memory frequency is 125 MHz, yielding a 2 GB/s bandwidth. Unfortunately, being rare as it is, I did not find any tool for overclocking experiments. Since this is a 128-bit card, SiS has left the low-end waters for more expensive mid-range. After heatsink removal, we can see the chip is indeed from early 1999.

Finally users got control panels for 3d settings. The palette really works only under Direct3d.

While the card itself is rather small, it drives multiple outputs thanks to the SiS 301 chip. The question you might be asking is why the heck does this card comes with a whopping 64 MB of memory? It looks great on the box, but you can also trust the maximum resolution. To my own surprise, even 3d rendering did not give up on 1920x1440. Sure, the framerate was impractical, but SiS laid down some solid foundations for the future. Video playback is also ready for the width of 1920, but this time the MPEG2 decoder is not very feature-rich, leaving most of the work to the CPU.

Experience

Since the card comes with a quality slider for 3d rendering, we need to settle on the setting. By default, the slider is in the middle position. I did not notice any particular difference by moving it to the max. On the other hand, reducing quality from default halved texture resolution along both axis and bilinear filter quality suffered as well. I decided to test the card with the highest setting because even then some worries about image quality remained. First, the texturing engine drops too many pixels from the furthest mipmap. This is a minor offense, hardly noticeable in games. The second worry is texture perspective correction. There were a few scenes in X: Beyond the Frontier that looked off and in Falcon 4.0 the ground was outright wobbly. Strangely these artifacts are this much isolated, so it is hard to say why they occur in these cases specifically.

The dizzying effect of shifting ground textures cannot be conveyed by a single screenshot.

Exceptionally, some specific textures are showing excessive color banding, probably because the driver chose the wrong format or additional dithering. When it happens to a wall texture in Thief 2, a player might not notice that until staring into it at zero distance. But when a texture as big as the sky in Unreal is severely banded the experience undeniably suffers.

Albeit visuals rendered are among the more pleasant. It is not yet as solid as the best of 1999 could do, but generationally speaking it is in the same league. The pipeline is designed for 32-bit precision and the horsepower for pushing such precision is there as well. All the Direct3D games I tried did run without any other issues. OpenGL is somewhat different, though. While dithering in 16-bit color depth under d3d was gentle, in Quake 3 a familiar pattern transpired. Combining transparent textures led to the usual accumulation of errors. Is there any remedy to this? In the case of OpenGL at least, I would recommend going with the 32-bit depth. This card can take it. One more thing without a cure was Sin benchmark. Some textures outright got bugged. You can see how exactly along with all the other screenshots in the gallery. On the other hand, Half-Life looked better under OpenGL so the measurements are from this API as well.

This kind of artifact on the ground textures of Battlezone was not observed on any other card.

To sum up the API support, it serves well under Direct3D, and OpenGL feels rather rudimentary. But remedies are available except for rare glitches. Outside of those few visual pitfalls, SiS300 was trouble-free - is a thing I would like to finish with. But my current system with the i815 board was weirdly unstable. Before more setups are tried, I shell withhold judgment.

Performance

While my benchmark suit is still mostly 640x480x16 it does not allow for more powerful cards to show disparity to the biggest extent. In comparison with the fastest 3d accelerator so far, the i752 IGP the SiS300 takes a clear yet not earth-shattering lead.

As usual, click to see minimal framerates.

This is an 18 % improvement in framerates. Not much, but it allows the SiS300 to take the lead in image quality as well. Because visual mishaps that did occur, can be mostly fixed by using 32-bit frame buffer and the performance impact is less than those 18 %. For Unreal the speed you see is already at true colors to get rid of that unfortunate sky texture. In case depth accuracy would appear insufficient, 32-bit Z-buffer is available as well and also with little performance setback. It is noteworthy how little drop can there be when upping the resolution from 640x480 to 1024x768, but vsync can skew the ratio.

Conclusion and decline of SiS

In comparison with the previous 6326 3d accelerators, the SiS300 is in one way a huge leap forward. However, we are talking about years when huge leaps were regular occurrences. Despite the performance achievements, the desktop market share conquered by the 6326 could be defended only by a solution that would offer something extra. And the 300 feature-wise looks only like the 6326AGP perfected. Be it a branding issue, higher costs, or driver quality, for some reason, the SiS300 was rarely sold. So rarely, I am afraid my stability issues might not be an isolated case.
In 1999 SiS remained among the technological leaders of integrated graphics chipsets, but in a time when it could not find any manufacturing capacity to actually deliver them in numbers the market demanded. SiS decided to go large and invested a fortune into its own manufacturing plants and also acquired x86 CPU technology by the purchase of Rise. The SiS300 would not go away without its offspring. Later die shrink and re-engineered with a 64-bit memory bus, the cheaper "305" sold a lot more, but still not enough to regain the former market share of 6326. SiS remained focused on low-end in 2001 and their first TnL chip "315" simply tried to do similar DX7 rendering as competition for "Asian" price. But the uncertainty of solid drivers kept retail cold. At least the chipsets business was again gaining momentum thanks to some great innovative designs. In 2002 SiS pumped out Xabre graphics cores and the long overdue holy grail of the company: complete system on a chip.

Peak of integration: SiS552 has x86 CPU, full featured north and south bridge, 2D & video pipeline of SiS315.

It is easy to understand why SiS introduced Xabre family with previously non-existing buzz, older cores got very little space in media. It again had a good price/performance ratio, was up to date in features, and was sufficiently fast in fixed function rendering. But drivers with low-quality texturing by default were cheap and Xabres just did not have enough hooks to turn the tides. With the graphics business going nowhere SiS decided to get rid of anything not essential for integrated graphics. The newly formed graphics-only XGI company released the second generation Xabre architecture under Volari name with more PR twists and hypes which fired back pretty heavily and confirmed the underdog position of XGI. After little more activity in the market, ATi purchased XGI in 2006. SiS was abandoning other challenging products through the years, sold its fabs and today does not offer any key PC core logic anymore, despite buying back the rest of XGI in 2010.