What's going on at S3?
Two years have passed since the DeltaChrome introduced S3's Columbia design. After CeBIT 2005 S3 has readied a Columbia refresh aka GammaChrome and recently announced a second refresh called Chrome 2x series. The purpose of this article is to show what has happened since then at S3 and where the aspiring 3rd power in the graphics industry is headed.
Current market situation
Before heading on to the technical aspects, we will take a look at the actual current sales situation of the DeltaChrome. For this purpose at CeBIT we spoke to Steffen Bruch, Club-3D's European Channel Manager. Steffen's task is to define which products may be marketed in Europe, and at what prices. According to him sales of XGI's Volari products, mainly the V3XT, are outpacing by far sales of S3's DeltaChrome series.
As reason of this he indicated that, although XGI's products are by no means superior to S3's, there is a broader palette of products, and with better availability than S3's offerings. Although on paper S3 may have interesting offerings for a number of markets, they are virtually not available. The targeted markets (entry level to lower mainstream) require, above all, cards at bargain and possibly incorportating a TV tuner.
Here S3's problems start. It has already been noted in the past that the time to market from presentation to delivery by far exceeds tolerable limits. The products simply arrive too late to be competitive. The DeltaChrome for example was announced in the beginning of 2003, and when it finally hit the shelves in Asia/US and Europe 1 year, or respectively 1 ½ years had passed. In order to remain competitive with its product, S3 had to sharply cut the selling price. The graphics cards market moves very fast after all.
One year ago GammaChrome, the refresh to DeltaChrome, was announced. Even in Asia you only have been able to buy this product since around June and in Europe it remains unavailable up to the current day. Even worse: This covers only the S18 with one quad pipe. The S19, which features two quad pipes, had been rescheduled to end 2005, and will now be replaced even before it appears by the Chrome 2x series.
There is even a potential danger that Crome 2x's mass availability may be delayed up to the point that, when it will finally hit the shelves, it will no longer be competitive. After all the competition already has fast mainstream chips which feature two quad pipes since more than one year. nVidia has demonstrated recently that even complex designs are feasible with older production processes and that the time to market may be decreased considerably. Maybe S3 should have used TSMC's 110nm process for S19 as well to ensure (at least for once) timely delivery.
Why are these chips with their good potential, such as the OmniChrome with its integrated TV tuner function, not available even for a card manufacturer like Club-3D? Why are, generally speaking, all interesting S3 products delayed? We believe that Fremont (California, S3's Headquarter) simply does not have sufficient liquidity at its disposal to order larger quantities of chips from a foundry.
Before you can make money you have to do an investment. This is a common truth. S3's parent VIA Technologies, which controls her subsidiary's cash flows, seems to be at the root of the problem. S3 is expected to function without investments, so there is only so much they can do.
The situation has actually even worsened to the extent that Club-3D is re-evaluating their commitment to further sell S3 products in Europe. As they are the sole provider of cards with these chips, this would spell the end of S3 products in Europe. Although Club-3D still advertises two products based on old S3 chips on its website, these are not listed in any shops at all. We seriously doubt that they will offer Chrome 2x based products at all, once the chip becomes available.
There is a different problem for S3: World-wide there are not enough card manufacturers which produce cards based on their products. Club-3D is the only important one. In order to solve this problem S3 is now manufacturing its own PCBs for the US market. This measure can, of course, only work as far as there are chips.
As long as the problems with the availability aren't resolved, there is no hope for S3 to go anywhere, despite their relatively fine designs, that much is for sure. With the next gen WGF2 chip code-named "Destination Films" things may start to improve. There are two development teams with a total of 350 engineers, new simulators and more know how. In our opinion the design should not be an issue here - let's hope that the production of chip will not fail due to a lack of available funding. After all, S3 needs to grab some market share to finally make the money they require to move the company forward.
Things are looking brighter in terms of end user support. For several months S3 have been running their Chromezone website where DeltaChrome and GammaChrome users can report errors and bugs, and there's also a list of compatible games available. S3 appears to also be cooperating more closely with game producers, so that DeltaChrome owners can now play nearly all available games. This has not always been the case.
Finally there's one aspect we need to complain about: Contact with the press. If you wish to sell graphics cards, the best and least expensive way to do so is to make the product known and appreciated by potential buyers via visibility in the press. S3 should step their press relations up, as the few reviews of their products tend to be done in the far east, while any form of western media coverage happens just every few months at best.
To summarize: S3 must quickly gain market share. This is only possible by offering a broad product range with sensible availability. In our view the designs themselves do fit the needs of the targeted markets nicely. S3 has a good chance to be successful with "Destination Films" if they get their act together beyond the chip design.
The technical progression
In a previous article we already highlighted the strong points and the weaknesses of the DeltaChrome design. In order not to make this article a prerequisite to understand the following, we would like summarize the characteristics in the following bullet points:
- bandwidth is scarce
- good raw fillrate
- moderate clock speeds
- average pixel shading performance
- sub-par vertex shader throughput
- angle independent anisotropic filtering
- supersampling AA with limited usability
- weak z/stencil throughput
- good early z/culling capacities
- unpartitioned and therefore inefficient memory controller
- low power consumption
- driver causes high CPU load
In the meantime a conversation coupled with a few investigations brought up some fresh architectural details of the Columbia architecture (we'll get there in a bit).
The DeltaChrome refresh named GammaChrome has already been announced a year ago. Contrary to our first thoughts that this would simply be a PCI Express port of the DeltaChrome, it is has become clear in the meantime that GammaChrome is a true refresh of his elder sibling.
The major change is at the raster stage: DeltaChrome's pixel pipelines are relatively inefficient. According to S3 both quad pipes can only work on the same triangle. This means rather simplistic control logic but it holds back performance in scenes with high poly counts (resulting in small average triangle areas). Consequently a large portion of DeltaChrome's fillrate and pixel shading resources goes unused.
nVidia and ATI are much more efficient in this regard. They use a much more sophisticated control logic. GeForce chips split the image into quad batches. A quad batch is a square with a side length of 32 pixels, with arbitrary orientation in the image. Radeon chips split the image into a fixed grid which consists of tiles with a size of 16x16. Then each quad pipe renders its part of the total number of tiles. What both solutions have in common is that each quad pipe has its own "private" are of pixels to work on, and there can be no interference by another quad pipe.
Rumor has it that this deficit has been removed in the GammaChrome design. Apparently the control logic has been modified in so far as that each quad pipe can now work on a separate triangle. The exact specifications are however unknown to us for the time being.
Pixel shading has been revamped as well. According to S3 Graphics the DeltaChrome pixel shader ALU can only perform FP16 operations in a single clock cycle, while FP24 ops require two cycles. We managed to verify this statement during our Shadermark tests. However, in game benchmarks we haven't seen anything that could benefit from the PP_ modifier. It seems like pixel shader throughput really isn't much of a bottleneck for DeltaChrome. With GammaChrome each pixel shader ALU is now full speed, which means that an FP24 op can now be executed in one clock cycle as well.
In a previous article we erroneously wrote that a texture operation would block the pixel shader ALU for a full clock cycle. We now believe that the DeltaChrome's texturing unit does not fetch the texture coordinates over the ALU's crossbars. Hence DeltaChrome is capable to perform one texture operation and one arithmetic operation in parallel per clock cycle.
GammaChrome's register count has also been bumped up, to raise the pipelines' efficiency. Unfortunately we do not have any details concerning the exact nature of the changes, but we strongly believe that the ALUs now support an increased number of special functions, or perhaps that a few "mini-ALUs" have been added. GammaChrome's shading power per pipeline and per clock cycle seems to have risen by a considerable amount.