by FastSite
02/02/2000 | 12:00 AM
If you have read our 3dfx Voodoo Scalable Architecture Preview, you probably remember that there were a couple of unanswered questions there. However, now we are ready to give you a few answers. We received some info from our own confidential sources and besides, our friends Dave Barron (from http://www.beyond3d.com) and Van Mardian shared with us the information they had got from the famous evangelist Scott Sellers who is known for his comments on 3dfx products as well as on some competing ones. We hope you do understand that for some ethical reasons we simply couldn't post any additions to our VSA Preview before our friends had done it.<%BANNER[article]%>
Well, before passing over directly to the questions and answers, we would like to make a few comments.
What is a preview, actually? As we see it, it is an article devoted to an already announced product, which hasn't appeared in the market yet. And can an article like that provide exact and credible information on the product considered? In fact, it seems to us almost impossible. The thing is that these articles are usually based only on the information available at the moment the preview is being written. And this info is usually provided by the manufacturer, which doesn't necessarily mean that it is complete and absolutely correct. Besides, different facts can be interpreted differently if the explanations are insufficient and unfinished. That's why if we have any questions or doubts we always mention them in such articles. As a rule, the actual truth comes to light only after we manage to test the real product. So the statements made in the previews may not always coincide with the real state of things. Does this mean that writing previews makes absolutely no sense? We don't think so, really. First of all, previews help us to get an idea of what we can expect in the near future. Second, if the facts mentioned in the preview prove fully correct, it will mean only one single thing: the manufacturer was absolutely honest with the users and kept his promises. Otherwise, the manufacturer's behavior can be regarded as misleading. Usually when writing a preview we interpret all the occurring doubts in favor of the new product manufacturer. Why so? Well, for instance because we will be able to express destructive criticism afterwards, which is usually more impressive, if some of our suppositions turn out wrong. Moreover, we all expect the coming products to be something extraordinary. The reality however, is much more down to earth.
We usually don't make any sequels to previews, because we never suffer from lack of information on the products previewed. And then in the product review we point out which of the promised things have been introduced in the end product. In case of VSA preview there were a lot of uncertainties. In fact, 3dfx didn't disclose anything on some product peculiarities, which caused a lot of questions and made us guess. We would like to stress once again that all the suppositions were made in favor of 3dfx. And now, when we finally found out that there had been made a few changes to VSA-100 chips and VSA technology, we decided to offer you this "second part".
And now let's return to VSA preview and take a look at some new things, which we happened to know.
The most interesting thing, which attracted most of you probably, was the possibility to use SLI and T-Buffer simultaneously. First of all, under "SLI" in case of VSA 3dfx understands something different from the old SLI technology used in Voodoo2 epoch. In other words, 3dfx simply used an old, already well-known name to denote a new technology. A new SLI version implies much more than ordinary frame lines interleave.
Today besides the traditional lines interleave technology they have also introduced bands interleave, where bands consist of several lines. Then SLI technology represents a communication protocol of 2-32 VSA-100 chips. In case the card is equipped with 2 or more VSA-100 chips, SLI will be used. Besides, band interleave may not be necessarily used in some cases.
Like we saw in the traditional SLI technology, the new one also uses the notion "Master chip". In the configurations composed of two and more VSA-100 chips one of the graphics processors is always a Master and all the rest are Slaves. The results provided by several chips (namely, the content of their frame buffers) are combined in a strictly defined order in Master chip frame buffer. Then from this buffer the data is taken to the monitor where it is displayed. Note that it is possible to combine separate pixels as well as the whole bands formed by different VSA-100 chips, if the chips work in bands interleave mode. As a result, we get a full frame, which is displayed on the monitor screen. However, the pixels forming the bands can be obtained from T-Buffer, i.e. when there are several pixel samples used for one pixel.
But controlling the display of the end frame on the monitor is not the only thing VSA-100 Master chip is responsible for. It also has to synchronize the entire work with Slave chips, which are busy distributing the data received from the system CPU. Sometimes, for example in case of a four-chip configuration in T-Buffer mode one of the Slave chips plays the part of a Master chip and synchronizes pixel samples combination process. We will dwell on it a bit later in our article.
Now we would like to make a very important comment before passing over to greater details.
3dfx described their VSA-100 chip as follows:
Taking into account everything mentioned above we could conclude that each of the two rendering pipelines in VSA-100 has two texturing blocks. There was not a hint implying that in multitexturing regime VSA-100 was capable of displaying only one pixel per clock. Of course, we were guided solely by the most righteous motives. However, almost right after the official announcement of VSA-100 and our preview, we got some info from 3dfx saying that VSA-100 could display only one pixel per clock in multitexturing regime. In fact, they resorted to the same solution, which has been used in NVIDIA chips since 1998, i.e. one rendering pipeline had one texturing block. On the one hand, this solution provides full pipeline utilization in all work regimes. On the other hand, in multitexturing, which is used almost in all the today's games, the theoretical fillrate reduces almost twice. It's pretty strange but 3dfx believe that such games as Quake3Arena do not actually use multitexturing the way they should. And to illustrate their statement they suggest taking a look at the sky, which is formed by three textures superposed in three passes. We would like to refrain from comments here, just note that if a game uses multitexturing it does not mean that it cannot also use polygons with one or ten superposed textures. The idea is that the more textures are laid over one pixel, the more time is actually spent which leads to the overall performance reduction of the graphics accelerator. Further on in order to show how it all works, we will consider only utmost cases, i.e. when the application always uses multitexturing or when two textures are always laid over one pixel. When some real graphics cards on VSA-100 appear, we will be able to arrange complex testing and to look at the real performance.
And now let's take a look at how two VSA-100 chips work in T-Buffer mode and without it. First, a picture of two VSA-100 chips on one and the same card:
If T-buffer is not used, a traditional SLI technology gets involved. Though this time it is slightly different:
And the difference is a new line bands interleave technology used besides the already familiar lines interleave. This band may contain from 1 to 128 lines of the frame and the number of lines in a band may change dynamically. Just in case you forgot: this ability to change the number of lines in a band allows distributing evenly the amount of polygon vertexes covered by one band, which provides even VSA-100 utilization. And using separate memory buses as well as using the available local graphics memory in a special way result into a noticeable performance gain.
Each VSA-100 chip has its own 128-bit memory bus, which allows distributing the workload evenly during data transfer and reducing the required memory bus bandwidth. The local graphics memory is used in the following way. The textures for each VSA-100 chip are stored separately and are identical, i.e. in fact the textures loaded into the local graphics memory are duplicated for each chip. All the other data, such as triangle vertex coordinates and depth (z), aren't copied and are single for both chips. They are located in the unified memory. And each of the VSA-100 chips forms its own frame buffer in this unified memory. Each of the two chips creates complete pixels in its frame buffer (to be more exact, rendering takes place in the back buffer). These pixels make up frame lines, and the latter form frame bands. Then these band are combined in the VSA-100 Master chip frame buffer. In the end the data is transferred to RAMDAC and then to the monitor and is displayed as pixels.
We feel like pointing out that in all color regimes the data will be transferred via two buses to two processors. However, each graphics processor will handle only a part of the transferred info, a half at the most. In fact this helps to reduce memory bus utilization and hence to increase the performance of the graphics card, particularly in 32-bit color depth. These advantages of SLI mode will have a positive influence in games requiring high fillrate at higher resolutions.
But if the user enables T-Buffer, the picture gets slightly changed.
To get a frame ready to be displayed on the monitor in T-Buffer mode several versions of this frame should be combined in a certain way. In other words, the technology uses several pixel samples or sub-pixels to form one pixel. Since all the T-Buffer effects are none other but some variations of full scene spatial and time anti-aliasing, we will take full scene anti-aliasing (FSAA) as an example, because this particular effect is absolutely transparent to applications and interfaces.
To create an anti-aliased pixel the technology uses several samples or sub-pixels, which carry the info on the color of the neighboring pixels. These pixels are chosen with a certain offset from the pixel waiting to be anti-aliased. Then the colors of the pixel samples are combined and we get the color of the anti-aliased pixel. And since in multitexturing regime VSA-100 can create only one pixel, 3dfx decided to use 1x2 mask for FSAA realization in dual VSA-100 configurations.
In other words, each anti-aliased pixel is formed by 2 pixel samples or 2 sub-pixels. However, in case of multitexturing two passes are required for each pixel sample or in the first pass two sub-pixels with a single texture appear and in the second pass the second texture is superposed. As for the most preferable method of forming pixel samples, there is no definite answer: the decision will be made on the run. As to 3dfx, they will choose the most effective way.
So, each of the two chips forms bands of anti-aliased pixels directly in their back frame buffers. Then these bands are combined in Master chip frame buffer into a full frame, and then the anti-aliased pixels are displayed on the monitor. We would like to draw your attention to the fact that this was the most general case. You should keep in mind that some variations are also possible.
For example, you can use 2x2 mask and hence there won't be any band interleave. Every 2 clocks both VSA-100 chips will form 2 pixel samples each and then the four pixel samples will be combined into one anti-aliased pixel. This will allow improving the FSAA quality, however, it will at the same time reduce the fillrate down to 83-92 million pixels per sec. That is why you will be ably to use this regime only at lower resolutions.
Of course, the user will have to decide himself if to choose this or that anti-aliasing method or to give up FSAA at all. But this is not all. In fact, you can find a compromise in case of FSAA and sacrifice the quality for the sake of a slight performance gain, namely, you can use 1x2 mask when forming an anti-aliased pixel. However, you will hardly manage to do the same thing with all other T-Buffer effects.
The thing is that such effects as Motion Blur or Depth of Field require at least four pixel samples or a combination of four slightly different frames to create one pixel or one end-frame. As a result, the user appears unable to utilize different T-Buffer effects except FSAA with 1x2 mask in dual VSA-100 chips configurations. Only if the game doesn't use multitexturing it may be possible to enjoy quite a satisfactory performance with T-Buffer effects enabled. And taking into account that such effects as Motion Blur should be necessarily supported by the application if you want to have them, they may turn out simply undemanded.
We doubt that right after the launching of Voodoo5 we will see any games using T-Buffer effects (except FSAA) or gaming patches providing the support of these effects. Why? Firstly, we haven't yet seen any announcements from game developers. Secondly, the situation with HW T&L is a good example of what may actually happen. Even under the most favorable circumstances only in three months after the first graphics cards on GeForce 256 had been sold we saw the first games supporting hardware T&L, which were very few, in fact. Besides, you should also take into consideration that NVIDIA spends a lot of money and effort promoting hardware T&L support among applications developers. If you remember, HW T&L is supported in DirectX 7.0 and in OpenGL. At the same time, T-Buffer, though it is announced as an open tool is still 3dfx's innovation and it is not so beloved among the software developers.
That is why 3dfx marketing managers on purpose separated FSAA from all the other T-Buffer effects and mentioned it as isolated in the list of Voodoo5 supported features. By the way, nothing prevents you from using combinations of different T-Buffer effects. For instance, you can combine Motion Blur and FSAA applying them to different parts of frames. It is possible because in both cases several pixel samples are required for a complete pixel and sub-pixels are combined on the hardware level.
So, let's find out how everything works in the most probable regime. Shall we say, we want to play at 1024x768 using FSAA with 1x2 mask. Then each of the two VSA-100 chips forms bands. Here we can say for sure that every chip forms only about a half of the whole frame. The lines form bands in separate (back) frame buffers and then the bands are combined in the Master chip front buffer. Then the data is sent to the monitor.
Note that the quality of this FSAA realization will be much worse than in case we used 2x2 mask. The difference will be most evident in inclined lines, where we sincerely expect to see no stepping effect. In other words, anti-aliasing will be insufficient and no straight inclined lines will please our eye. Besides, the theoretical fillrate will drop in multitexturing regime. For example, in multitexturing using FSAA with 1x2 mask the fillrate will make about 166-183million pixels per second. This should be more than enough for playing at 1024x768 in 32-bit color and at 60fps. Well, we'll see what the real tests show.
And now we suggest looking at four-chip configuration working with and without T-Buffer. First comes the picture showing four VSA-100 chips on one card:
We find it important to mention that depending on the working regime this configuration can be built with one Master chip responsible for the synchronization of the other three chips. Or there may also be one more Master/Slave chip, which obeys Master chip, on the one hand and coordinates the work of the second pair of chips with T-Buffer enabled, on the other hand.
When T-Buffer is disabled, the situation hardly differs from that with two VSA-100 chips.
Each VSA-100 chip unites frame lines into bands in its frame buffer (back buffer) and then the data from the back frame buffers of Slave chips is transferred into Master chip frame buffer where it is combined in a certain way. After that the data from the Master chip frame buffer is sent to the monitor. Since the data transfer between the frame buffers takes place inside the unified memory everything happens very quickly and shouldn't have any negative influence on the performance. The peak fillrate in this case may reach 664-732 million pixels per second in multitexturing. If the user decides to enable T-Buffer, for instance FSAA, then four VSA-100 chips will work in the following way:
Every chip pair deals with forming lines of anti-aliased pixels, i.e. in fact, every chip pair creates only a half frame. For FSAA with 2x2 mask, which provides really good anti-aliasing, each anti-aliased pixel is formed of 4 samples or sub-pixels. In this case with multitexturing enabled each VSA-100 pair needs two passes to form an anti-aliased pixel, because two chips can form two pixel samples per clock. By the way, it will probably be possible to use FSAA effect with 1x2 mask, so the user will have a choice.
Without multitexturing an anti-aliased pixel is formed during one clock. Since each VSA-100 pair forms only a part of a frame, the workload is distributed evenly and the fillrate drop in multitexturing doesn't affect the fps that much. Even in multitexturing with FSAA effect enabled the theoretical peak fillrate will be about 166-183 million pixels per second, which is quite a lot for today. Just for you to compare: a graphics card based on GeForce 256 working at nominal frequencies in multitexturing but without FSAA achieves peak fillrate of 240 million pixels per second.
Keeping in mind everything mentioned above concerning the work of T-Buffer in dual-chip configuration, we have every right to state that only Voodoo5 graphics cards with four VSA-100 chips on board (Voodoo5 6000) will allow using all T-Buffer effects of acceptable quality and at a quite satisfactory speed. Why are we speaking about acceptable quality, you wonder? No matter that a combination of 4 pixel samples is enough for such effects as Motion Blur, the combination of 8 pixel samples (or 8 frames) or more will lead to a definitely better result.
We doubt if anyone will wish to play at lower resolutions with FSAA effect, because the image will be evidently too blurred, which can surely make not the best impression. In fact, the optimal resolution for any game using FSAA is 1024x768 especially since the fillrate of Voodoo5 should be just enough for playing at such resolutions. However, if the user disables T-Buffer, he will be able to feel very comfortable when playing at the resolutions up to 1600x1200 at 60fps. Besides, at higher resolutions the detailization is also higher that is why FSAA is no longer so important there. And what will an average user prefer? Hard to tell. We can only remind you that most ordinary gamers are still playing at resolutions up to 1024x768. Of course, some users can boast larger monitors. Probably some of you can even afford to buy a graphics card for 600 bucks, which will allow using all T-Buffer effects at an acceptable level of performance. But we really doubt that there will be many users like that. The owners of such monitors, the so called hardcore gamers, will most probably prefer high resolution to lower resolution with FSAA. As a result, with a graphics card for $600 they will never use all its features. In other words, Voodoo5 seems to have a very gloomy future.
And now a bit more on VSA-100 details.
VSA-100 specification told about hardware bump mapping. And in fact, it turned out that the chips support only Embossing, which is the least efficient of all. No Dot Product, not to mention EMBM.
Besides, the specs also mention full AGP 4x support. In fact, under full support 3dfx understands supporting everything except such a trifling thing as DME or AGP texturing as well as Fast Writes mode. We were happy to find out that at least SBA was supported. 3dfx explained the absence of DME support with the lack of games supporting this mode. Now 3dfx isn't going to support this mode in its future products. If the game needs to load huge textures, then the game should be using texture compression, according to 3dfx. Of course, this position does make some sense, however we still can't make out what prevents 3dfx from introducing DME support? Does it have to do with any technical problems?
Everybody was very curious about the way 3dfx was going to make its products work in 32-bit color depth. They wondered if there would be any post-filter used for that. Today we know that there won't be any post-filter used for 32-bit color regime. And in fact, it is very good, because firstly, it will save 3dfx time and trouble replying to all the reproaches and attacks and secondly, a standard and correctly carried out solution will be more than enough. This is 3dfx's first experience in dealing with 32-bit color that is why they should better refrain from making too many experiments. The post-filter will be used only during the rendering in 16-bit color. The post-filter in this case will be the same as in Voodoo3.
Now we would like to comment on the question about so many transistors used in VSA-100 chip. 3dfx motivates it with a complex hardware realization of T-Buffer and an increase in texture cache size. However, the texture cache size is not stated, which has already become typical of all the manufacturers. We can suppose that the texture cache of VSA-100 is of 32KB.
Note that in case of the configurations with two (and more) VSA-100 chips, some chip functional parts may remain unused. For instance, each VSA-100 chip has an integrated RAMDAC, however, only the RAMDAC of Master chip is actually used. The same thing happens to the block responsible for hardware T-Buffer support, namely for pixel samples combining. Why didn't 3dfx use one external RAMDAC and a separate chip responsible for hardware T-Buffer support? For purely economical reasons. It is cheaper to duplicate some functions in all the chips than to spend money on the production of separate chips with some special features. The same thing refers to RAMDAC - it is much cheaper to integrate it into all the chips than to use one external RAMDAC.
In 2D graphics works only one VSA-100 chip. And all the other chips are not involved. Actually, as far as we know, 3dfx plans to install a second VGA out onto Voodoo5 cards. In this case there will be two chips working in 2D graphics. If we consider a four-chip configuration it will be possible to display 3D graphics on two monitors.
As far as the chip with Intel marking used on Voodoo5 6000 is concerned, we failed to find out anything new here. We only happened to know that Intel chip (it is seems to be something like non-transparent PCI-to-PCI bridge, but not necessarily) is required to allow giving up power supply via AGP port completely and at the same time to maintain normal operation of the card in the system.
Concerning the reasons that guided 3dfx when they made up their mind to produce cards with PCI interface. Here everything is very simple. Intel is the one to blame (we can't help adding: as usual). Why Intel? Well, because now Intel doesn't provide anything at a reasonable price instead of a slightly outdated i440BX chipset, besides some i810 variations. As you know, i440BX doesn't support AGP 4x, UDMA/66, PC133, i.e. those chipset functional features, which appeared necessary (as a result of successful promotion). i820 chipset is intended for very expensive DR DRAM, and i815 (Solano) so impatiently awaited by numerous users may be postponed for an indefinite period of time. So, the user is forced to buy mainboards on i810, which means no AGP port, or on VIA Apollo Pro 133A. If you buy an i810 based board you won't be able to connect an external graphics accelerator with AGP interface. This is where Voodoo5 5000 family may turn out in the right place. Nevertheless, it is still very hard to say if 3dfx managed to predict correctly the future of the graphics market, because there are 2-3 more months left till the first Voodoo cards appear.
Let us give a brief commentary on Voodoo5 prices. 3dfx representatives stated that Voodoo5 graphics cards will be equipped with SDR SDRAM, because this is the cheapest memory in the market right now. However, they explained very high cost of Voodoo5 again with the memory prices. But there is one very interesting detail. Everybody knows that SDRAM has become more expensive during the last 3-4 months. However, the prices rose only in the open market. In fact, 3dfx is very unlikely to be buying the memory in the open market since it should definitely have some long-term agreements. The contract memory prices have hardly changed, at least they didn't get several times higher. So, 3dfx seems to be slightly cunning so that to be able to maneuver just in case. That is why we have every reason to suppose that by the time Voodoo5 cards are launched their prices will be a bit lower than those we have now.
Not so long ago 3dfx announced that their own closed Glide interface was becoming open for everybody. Unfortunately, we can't say for sure how 3dfx can benefit by it. We don't think that applications developers will return to Glide. Firstly, if 3dfx competitors don't support API Glide in the drivers for their products, then the developers won't have any reasons for supporting a not very popular interface though open. Secondly, the shadow of Microsoft is constantly hanging over the software developers, and this company is very unlikely to favor the spreading of Glide. Thirdly, Glide interface is strictly aimed at 3dfx chips and their features, the functions it supports are limited by the functions of 3dfx Voodoo chips. And if we compare the amount of functions supported by Glide to that supported by OpenGL, then Glide will appear a very specific OpenGL subset though there is no direct connection between these two API. It will be more correct to say that they are similar but not compatible. We would like to note that API Glide has low-level functions absent by OpenGL, such as texture memory operation (grTexMinAddress, etc.). OpenGL doesn't allow controlling texture memory distribution. If we take possible combinations of pixel color regimes, such as texture*alpha, again Glide will offer us a few regimes, which cannot be reached from OpenGL in one pass, for instance iterated rgb + texture. Both OpenGL and Glide interfaces support quite a lot of common functions. However, we should mention that OpenGL belongs to a much higher API level than Glide and allows describing the tasks in terms of light sources, meshes, positions and turns, while Glide is a sort of Voodoo wrapper and in fact simplifies the accelerator programming at the lowest level. We can probably call glide an OpenGL subset only meaning that you can write for OpenGL the same way you write for Glide - without integrated into OpenGL T&L, etc. Besides, the similar ideology of both API is also worth mentioning: most functions carry the same idea and differ only by prefixes (gl by OpenGL and gr by Glide) and a bit by names such as glBlendFunc/grAlphaBlendFunc etc. Each new Glide version makes it closer to OpenGL: for example, in Glide 3.0 grSstIdle () was replaced by grFinish() (the functions are very close to glFinish()). However, as it comes to vertex representation, Glide and OpenGL can hardly boast any considerable similarities. Glide is more like Direct3D with its flexible vertex format. Besides, you should also bear in mind that quite a lot of developers use OpenGL and they won't suddenly turn very anxious to shift to Glide. Especially since OpenGL has already become multi-platform and ported. In other words, Glide may probably get widely spread only in operation systems such as Linux, which will allow playing some games there too. However, if the game uses Direct3D, then Glide support will hardly have any effect in Linux. Nevertheless, the situation is still very interesting to watch.
And now let's take a look at the future of T-Buffer effects in real life. If you disregard T-Buffer effects and hardware texture compression, then Voodoo5 graphics cards will turn out a sort of variation of competing products such as NVIDIA GeForce 256 but with a much higher fillrate and without hardware T&L. In fact, the user will have to choose whether to play at high resolutions and at high fps rate or to play at resolutions up to 1024x768 at 60fps and with only one T-Buffer effect - FSAA. All the other T-Buffer effects are not transparent for applications and interfaces. It means that you will be able to use these effects only if the game developers want it. T-Buffer may win developers' acknowledgement a bit quicker if it is supported in such standard interfaces as Direct3D. Now we know only that 3dfx negotiates this with Microsoft. But even if we make the most optimistic forecasts, this support may appear only by autumn 2000 when we see a new generation of 3dfx chips. And now they are trying to make us buy what we won't ever use.
What will the user's choice look like? It's a difficult question. Since there is no definite answer, the advantages provided by T-Buffer start raising doubts. As for hardware texture compression, there turns up a problem with support of this functional feature by applications. No wide support is planned now. As a result, the commercial attractiveness of Voodoo5 graphics cards loses quite a lot. And now add here serious losses 3dfx suffered in 1999, the announcements of very expensive products and unclear situation with the sales volumes.
Since most of you are very fond of various forecasts, then you should try to imagine the market of March 2000. 3dfx will start selling its new graphics cards featuring the things, which have been promoted by its competitors since the end of 1998. Besides, these products will also possess a number of absolutely unique features such as T-Buffer, hardware texture compression and good scalability without any clear future, however. In fact, 3dfx is late with its products. If Voodoo5 graphics cards had appeared in autumn 1999 as everybody had expected (and 3dfx competitors as well), then 3dfx could have remained among the market leaders. But somebody from 3dfx has made an inexcusable mistake, and now even very clever Scott Sellers fails to convince the public that 3dfx has no problems at the moment and all our fears and concerns are absolutely ungrounded. As for us, we wouldn't buy 3dfx shares right now.
So, let's imagine that we are in March of the year 2000. 3dfx is proudly introducing its new products. At the same time NVIDIA launches its new GPU GeForce 256 II (NV15) with 600-800 million pixels per second in multitexturing and HW T&L. NVIDIA may even turn out an indisputable leader at least in terms of performance. And if we suppose just hypothetically that there won't be any other promising solutions by that time, which will be able to compete with NVIDIA and S3 won't solve the problems with Savage2000, then NVIDIA undoubtedly becomes a new monopolist. Besides, no one can guarantee that NVIDIA will use its monopoly for everybody's good. For us it may result into higher prices on NVIDIA products, for instance. While 3dfx is slowly declining…
Of course, we considered the most pessimistic situation and laid on the colors a bit too thickly. But this case shouldn't be absolutely disregarded, especially if 3dfx doesn't undertake something the sooner the better and tell about its future plans. Anyway, we'd better imagine the worst state of things now so that we could be happy if something turns out better than we have imagined.
And now let's sum up.
Well, what can we say? We think that 3dfx is sinking in problems only because of the unskilled managing. It is mainly the heads of the company that make strategic decisions. In our opinion, 3dfx chose the wrong policy in 1998. As a result in 2000 we will have the whole bunch of products at higher prices and with very doubtful advantage compared to the competing products.
Of course, if we were more objective, we would say that 3dfx lost too much on marketing, in the first place. Its competitors managed to convince everybody that 3dfx products didn't have the functions required at this particular moment. So, 3dfx is defeated. At least now.
It looks as if the 3dfx board of directors has finally realized that the managing core should be changed and has already done it. We can't predict if the new president proves able to save the situation. At least it seems a clever move to reshuffle the marketing department. Now we have every reason to say that the first half of 2000 will be the hardest for 3dfx in terms of technology as well as finance. We think that 3dfx will manage to stay in the market only due to low prices, correct attitude towards its customers, clever advertising policy and serious economizing. We hope new 3dfx administration will be able to overcome these hardships and very soon we will see with our own eyes the announcement of VSA-200 featuring four pixels rendering per clock and T&L hardware support together with T-Buffer and texture compression.