X-bit labs - Print version

On 31 August at 6 p.m. Moscow time the world was about to change. Everybody was looking forward to this extraordinaryevent, which should expectedly take place after the official announcement of NVIDIA's new graphics processor. Themysterious parameters of this novelty had been rumored in and out long before the Day came. We also never disregardedthis topic, and the time proved most of our suppositions. To tell the truth, NVIDIA managed to impress the public withthe way it launched into the world its new chip called GeForce 256. It definitely deserved a particular place in the worldhistory. The efforts of NVIDIA's marketing and advertising departments were not vain. The whole occasion was very accurately,thoroughly and efficiently arranged. First, we would like to point out that NVIDIA did its best to prevent all possible leaksof info about the new graphics chip (further we will refer to it as to GPU - graphics processing unit, because it was NVIDIA,which offered this particular abbreviation). NVIDIA's officials bluntly refused to disclose to public any facts about theirnewly born child. On the other hand, such steady silence of NVIDIA gave way to numerous suppositions and forecasts of hungryanalysts, which also suited NVIDIA's plans and contributed to the growing tension and shade of mystery around this chip. Thepublic interest and curiosity kept increasing and from time to time NVIDIA just warmed up the crowd easily, hands down. NVIDIAprovided its main partners with the GPU samples in advance. Among those happy ones, who managed to get it, were a lot ofgraphics adapters manufacturers and even PC manufacturers. This move allowed NVIDIA to make a part of the computer industryexpress their highly positive (only positive!) opinion about the chip in advance. 18 hours before the official announcementof GeForce 256 GPU we all saw a very promising message on NVIDIA web-site: "In 18 hours the world will change...". You shouldagree that only this single phrase is worth your loud applause: a really strong and fine start! And the desired effect wasmade: all of us, who expressed at least the slightest interest to the coming event, were impatiently staring into themonitors watching the countdown and praying for the time to pass quicker. Something supernatural was about to happen. Theinvisible threshold was expected to be finally crossed. Time passed... but, god, it passed so slowly!<%BANNER[article]%>

Yes! Finally! Now! ... but NVIDIA's server working under Lotus Domino appeared unable to cope with all the http requests.The whole thing ended up in great puzzlement and even despair of those, who failed to be there when "it" happened andcouldn't find out what, in fact, had changed, because they simply couldn't get to NVIDIA site. However, other sitesdealing with computer news and hardware stuff immediately stilled the informational hunger of the public with numerousdetailed reports on what had really happened. Had the world changed? In some way, yes. But though there was so much fussaround GeForce 256, this event can hardly be called a revolution. It is more likely to have turned out the next stage ofcomputer evolution, that's it.

The incredible enthusiasm of the excited public wasn't spoilt even by the announcement of S3 Savage2000 the day before,which was supposed to be as cool as GeForce according to the declared parameters. However, NVIDIA preferred to pretend thatthere hadn't been any announcements from S3 and kept on defining GeForce 256 as "the first graphics processor with theintegrated geometric accelerator intended for the mass market". But Savage2000 from S3 still remains the first officiallyannounced product of this kind despite all the statements and definitions like that. And as for the silly mistake inGeForce 256 specification list stating that the vertical refresh rate was 75MHz(!) at the resolutions up to 2048x1536,it remained absolutely untouched by the mass media. Remember the auction at eBuy where the first graphics card based onGeForce 256 signed by its developers was sold for $8100, while the starting price was only $290? Well, we should dojustice to NVIDIA, it really deserves our respect and admiration. The recent situation in the world proved that theirproducts are commercially successful and enjoy wide popularity due to their perfect technical characteristics andhigh-quality drivers.

So, NVIDIA called its new product GeForce 256. Is it a good name? Well, before we try to answer this question wesuggest recalling that NVIDIA arranged a sort of a contest for the best name. Tastes differ, you know. Somebody may becrazily happy about the name and some of you may consider it awful. Anyway, pretty soon we'll get used to this name andwon't care about any comparisons, which may occur to us. For example, the first idea that rushed through our mind wasabout G-Shock or G-Man. However, if we refrain from different sarcastic comments, we will easily understand the aim ofNVIDIA marketing specialists: the name was supposed to imply the main features of a new graphics processor. Let's tryto single out some meaningful parts. Ge undoubtedly means that the device is equipped with the geometric coprocessor(geometry - Ge). Force symbolizes the computing capacities, and 256 indicates that the chip and the memory bus have256bit interface. And now let's take a closer look at the parameters list of GPU GeForce 256.

General Characteristics:

2D Acceleration Features

3D Acceleration Features

DVD/DTV Support

OS, Interfaces and Software Support

Here we would like to note that any suppositions about GeForce 256 being a combination of two TNT2 cores with a geometricaccelerator and higher bits are absolutely false and are none other than suppositions. NVIDIA's engineers developed GeForce256 graphics processor having started the work from the very beginning, and the whole thing took not 6 months but at leastover a year. We won't dwell on 2D features, because there can hardly appear anything new in this field. Of course,something can be spoilt but we believe NVIDIA has no problems of the kind judging by the perfect quality of TNT2 based 2Dgraphics cards.

Note that GeForce 256 has 23 million transistors and is made with 0.22 micron technology. It looks as if NVIDIA failedto finish all the preparations for complete shift to 0.18 micron technology. We think that NVIDIA had to limit GeForce256 core clock frequency by 120MHz only because of the 0.22 micron technology and greater number of transistors.Otherwise, there could appear some tangible problems with chip overheating and even with power supply. Accordingto the info we have at our disposal right now, NVIDIA also considered the possibility to equip its chip with aspecial air funnel, which was supposed to remove warm air outside. Judging by what we got in the end, this ideawon no support among the graphics cards manufacturers. And as a result the core frequency is not so impressivelyhigh as it could be according to the present day situation in science and technology. Besides, some graphics cardsmanufacturers, such as Leadtek, for instance, are going to integrate a special temperature monitoring system in theirGeForce 256 based products. Another cards manufacturer, Elsa, is planning to reexamine the reference design of GeForce256 based graphics cards saying that they have to find a better solution to the problem of overheating and power supply,which will undoubtedly delay the launching of their graphics cards for about a month. In other words, we should be readyto face real overheating problems, and hence, we will have to forget about overclocking once and for all, if the situationproves that serious.

Now let's particularly dwell on a few most interesting peculiarities of GeForce 256

The first thing, which catches your eye, is, of course, the integrated geometric processor dealing with coordinatestransformation, lighting and polygons sorting. We won't bore you with detailed discussions about the geometric processorand its main functions, because we have already touched upon this problem in our Savage2000 Preview. Here we are goingjust to mention the key issues about the geometric coprocessor. The geometric part of GeForce 256 responsible for T&Land Clipping boasts a really impressive computing capacity. Just imagine: GeForce 256 can provide 15 million texturedpolygons with Z-buffer every second! It's a very significant amount.

At the same time the system CPU can be absolutely free from transformation and lighting operations, which usually takeabout 70-90% of the CPU resources. In this respect there is a question: is there any serious need in graphics cards withthe GPU right now? As to the applications developers, in a year or so there will hardly be any games left working withouta geometric accelerator in the graphics card. They state that the difference between the performance of the graphicsaccelerator with the geometric coprocessor and the one without it is just incredible, which leads to considerabledifference in image quality.

As for the practical importance of the hardware vertex coordinates transformation with the help of a geometriccoprocessor, the thing seems quite clear. All OpenGL and Direct3D games (slightly modified) can make real use of thisfeature. However, as it comes to lighting, some problems still arise. The first lighting source is calculated by thegeometric coprocessor integrated into GeForce 256 "for free", i.e. leading to no performance reduction, while all therest (which is 8 light sources altogether) reduce the overall performance by 5-7% at the most (as to the preliminarycalculations). The game developers consider vertex lighting to be carried out too stencil-like. For example, there weresome problems with creating realistic candle shimmering: they failed to randomly change the brightness for differentvertexes, and if the brightness was changed for the whole source the effect appeared artificial, unrealistic. But thetechnologies and tools keep developing, and besides there is still enough time because NVIDIA predicts vertex lightingto be widely spread only by the end of next year. So, let's wait a bit more...

Then, GeForce 256 can boast 4-pipeline architecture. It means that every time step there can appear 4 pixelssimultaneously. Since the GPU clock frequency is 120MHz, the fillrate will make 480 million pixels per second. However,the pipelines of our hero have one texturing block each, which means that at the same clock frequency equal to 120MHz thefillrate will drop down to 240 million pixels in applications using multitexturing (when two textures are superposed overone pixel). Note that only filtered textures are superposed and 8 texture samples are used for each texel, which provides"free" anisotropic filtering. To tell the truth, anisotropic filtering based on 8 texture samples is quite an interestingsolution, however, the samples seem to be insufficient. It could be much better to involve 16 or even 32 texture samplesto define the final texture color. In this case the final color definition requires 8 samples, which should in the endprovide high quality image. Why does NVIDIA use only 8 samples and not more? Because this realization can be carried outwith one texturing block only, which was made in GeForce 256. The samples are selected according to the anisotropicdependence on the detailization along X- and Y-axes of the screen (LOD_X, LOD_Y). The sampling field on the texturelooks similar to an ellipse with the ratio LOD_X/LOD_Y. Gauss method very well suits as a sampling function. The samplesare taken from one and the same mip-level. If the ratio is not big the image will be deprived of blurring and washed outeffects, however, if the ratio is quite significant, 8 samples won't be enough, and some artifacts will turn up.

What does this all mean? It means that the quality of the image you see on your monitor will become significantlybetter. We have every reason to believe that you will be able to notice the change even with the naked eye, comparedfor instance with TNT2 based graphics cards.

And now let's say a few words about fillrate. This is the main factor, which influences the performance of anygraphics adapter. So, after GeForce 256 announcement a lot of editions expressed their concern that low graphics corefrequency would lead to low gaming performance. A bit later NVIDIA's competitors represented by CTO 3dfx Scott Sellersalso shared their opinion on the matter. In particular, Scott said that GeForce's fillrate is too low and insufficientfor the card to be used in today's games at high resolutions. Of course, he didn't forget to mention that the productfrom 3dfx would boast a much higher fillrate and would allow playing games at high resolutions without any performancelosses. In fact, according to the numerous rumors about GeForce 256, its fillrate was allegedly promised to be 3-4 timeshigher than that of TNT2. However, it never happened. For instance some cards based on TNT2 Ultra can boast the fillrateof 190mln pixels per second with two textures superposed over one pixel (in games with multitexturing). As for GeForce 256,its potential fillrate is 240mln pixels per second with two texels superposed over a pixel per time step. Well, let's studythe matter more attentively. We will try to find out the required fillrate for a particular resolution and the desired fps.Let's take the most optimal resolution for today: 1024x768 (we consider it to be most optimal because most users have 15"and 17" monitors). In order to get the number of pixels needed to be displayed on the screen to form the scene, we have tomultiply 1024 by 768. It makes 786,432 pixels for a scene. Now multiply this number by the desired frame rate. As far as weknow, the most desirable frame rate today is 60fps. So, 786,432 x 60 = 47,185,920 pixels per second. Seems quite a smallnumber, doesn't it? But we have to take note of another very important factor. The thing is that in case of traditional,the so-called polygonal, 3D graphics display architecture the scene is rendered successively - the textured triangles aredisplayed on the screen one by one. In fact, the scene is never known in advance, and the location of the scene objectdepends on the depth values stored in Z-buffer. Here is an example to illustrate our complicated ideas.

Suppose we have to create a scene with a car standing in front of the building with a girl sitting on its hood. (Whatshe is doing there, whom she is waiting for and why she is alone has nothing to do with our discussion). The traditional(polygonal) graphics accelerator will render the scene as follows: at first it displays the polygons with texturesforming the wall of the building, then those forming the car and then those forming the girl. The info about the depthof the objects involved in the scene stored in the Z-buffer allows defining which polygon should be on top and whichunderneath. Note that in our example the scene depth equals to 3, i.e. it is the maximum number of superpositions ofsome objects over another ones. All polygons of the car cover some polygons of the building wall, and all polygons ofthe girl cover some of the building and some of the car. So, it turns out that the graphics processor does extra work,which is absolutely useless: it could have saved time and trouble by simply leaving out those pixels of the buildingwall, which were covered by those of the car, because they are not seen anyway. And so on and so forth. As a result,we come to the definition of a new notion - Depth Complexity. Depth complexity shows how many objects are covered bya particular scene object. That's why if we return to our example the depth complexity of the building wall will beequal to 1, that of the car - 2, and that of the pretty girl - 3. Depth complexity is directly connected with overdrawvalue, which shows how often we need to draw a particular pixel anew. This happens if some pixels taken from the polygonsof the car cover those of the building wall, etc. So, overdraw defines how many times we will need to redefine the pixel,which will get to the monitor from the frame buffer. It's clear that the more complicated is the scene, the higher isoverdraw value for each pixel of this scene. This overdraw value for each pixel of the scene allows defining screenoverdraw, which is the average overdraw for the whole scene. The general formula for overdraw looks as follows:

[depth complexity of object N₁ x number of pixels in polygons forming the object + depth complexity of objectN₂ x number of pixels in polygons forming the object + �+ depth complexity of object N_m x number of pixels in polygonsforming the object] / [total number of pixels in the scene x 3] / [scene depth x total number of pixels in the scene] =overdraw (or screen overdraw).

[1 x number of pixels forming the wall of the building + 2 x number of pixels forming the car + 3 x number of pixelsforming the girl] / [total number of pixels in the scene x 3] = overdraw

Of course, the final value may not necessarily be an integer. Besides, even the formula we gave can be simplified a bit.In order to get the estimation value the formula undergoes the following simplifications:

[depth complexity of N₁ + depth complexity of N₂ + � + depth complexity of N_m ] / scene depth = overdraw

In our example some pixels are redrawn once, some - twice and some - three times. The scene depth equals 3. So, weget: [1+2+3] / 3 = 2.

Then if we display our scene with the girl and the car in front of the building in the resolution 1024x768 at 60fps,the required fillrate will be equal to 47,185,920 x 2 = 94,371,840 pixels per second. In practice there is no tool tocalculate scene overdraw for each particular application, however, the games of 1999 are considered to have the averagescene overdraw of 3 or 4, or even 5. In Quake3Arena for instance, this value is supposed to be 3. Then in case ofQuake3Arena run at the resolution 1024x768 at 60fps and with screen overdraw equal to 3, the required fillrate willappear 141,557,760 pixels/sec. In fact, there are some facts proving that the real overdraw in Quake3Arena is only 2.3.Note that the whole static world in Q3A has an overdraw value equal to 1, i.e. zero overlapping, achieved thanks to aspecial data format - BSP (Binary Space Partition). This format allows using all thoroughly sorted polygons, which formthe static world. Hence z-fill is carried out without the preceding read/compare operations, which deprives of thenecessity to separately clear z-buffer. For portals there is a special Portal Rendering Algorithm, which providesoverdraw equal to 1. Only a thoroughly built model of the 3D world provides this result. After the static world iscreated it is covered by the moving objects, such as the players for instance. Mirror surfaces are generated throughtwo renderings with the help of the stencil. But it is not Q3A that we are discussing here, so let's continue. If wetake for instance TNT2 Ultra based graphics cards with the graphics core clock frequency of 190MHz, they will performat about 60-70fps, which will be another proof to the theory. This fact states, by the way, a really high quality ofOpenGL ICD drivers from NVIDIA and clever OpenGL interface realization. But if we take D3D game Expendable, the fps willstop somewhere around 50. This happens because it is not only scene overdraw that influences the fillrate. There areanother two very important factors: the quality of the math1ematics operating the graphics processor (read: graphics carddriver) and API optimization level.

As far as the dependence on the driver quality is concerned, practical use proves this point quite excessively,when a new driver causes the fps increase though the system remains the same. Frankly speaking, NVIDIA seemed to haveno problems with the drivers recently. Moreover, not so long ago we happened to know that a team of SGI specialistsjoined NVIDIA, which for sure guarantees high level of optimization and high quality of all the software in the future.As for the influence of API on the performance, this question is pretty complicated, however, most applicationsdevelopers don't deny that if it were not for API calls optimization of the work, the performance drop would beinevitable. If you take a look at Direct3D listings, for instance, you will be able to notice the whole bunch ofredundant security measures, even if you are not a software expert. They constantly check if there are no negativevalues, etc. So, the conclusion will sound as follows: the graphics cards based on GeForce 256 will provide higherperformance in all today's games compared to the performance shown by the cards on TNT2. Turning back to the remarkmade by Scott Sellers, we would like to remind you of his idea to consider 1600x1200 the best example stating thatGeForce 256 will lack the appropriate fillrate to provide acceptable gaming. What can we say in response? He is right,GeForce will fail here. But in this case we also feel like asking: do many people play at such a resolution?

Just for you to know, in non-traditional (tile) architecture introduced in PowerVR250 the scene complexity (overdraw)value is always equal to 1, because the whole scene is defined before it is rendered.

As for overdraw, we also have to point out one issue. This value greatly depends on the polygons sorting order when thescene is created. To make overdraw value smaller, the polygons should be displayed starting from the nearest ones and thenpassing to the farthest. And to correctly process alpha-channel, in other words, to blend, for better image quality of thegame scenes, the polygons should be displayed starting from the farther ones, i.e. vice versa. To achieve higher performancethe polygons need to be sorted according to the textures used and to the vertexes order. Overdraw value can become considerablysmaller if we disregard those polygons, which are completely covered, from the very beginning. However, it is the task for gamedevelopers, who don't often feel like devoting it enough time and effort. By the way, overdraw will reach its maximumvalue if multitexturing is disabled. And one more thing: it is almost impossible to guess the overdraw value by sightnot knowing the sorting order, so don't even try.

Here we once again return to the problem of insufficient fillrate. The thing is that if overdraw gets down, the imagequality will automatically become worse, while if it gets up on the contrary, it will require high fillrate. However, wehave already proven that GeForce 256 can put up a good show, because its fillrate is more than enough for one of the mostpopular resolutions nowadays - 1024x768. Besides, high fillrate allows using different special effects improving the imagequality and hence adding more realism to it. We have already mentioned full-scene anti-aliasing, but there is also such aneffect as bump mapping. So since we came to speak about bump mapping, we would like to draw your attention to the fact thatGeForce 256 provides hardware support for two techniques: embossing and dot product. There can hardly be anything surprisingin embossing, while dot product is something pretty vague and unclear. No doubt, GeForce 256 does support Dot Product BumpMapping. However, it still has one drawback: the interpolation between the normals is linear instead of spherical, whichcan cause certain artifacts - when a lot of pixels on the screen correspond to one and the same texel in the texture withnormals. Now you see the difference between Dot Product developed by NVIDIA and the so-called traditional method presentin PowerVR250, for instance.

But we have veered from our main discussion point - the performance. Let's consider a few other important things here.

You won't deny that the card's performance also depends on such factors as additional special effects. In particular,full-scene anti-aliasing is expected to cause fillrate drop and hence lower fps in games. It is very likely to be so,actually. But NVIDIA stakes another GeForce's feature, which is supposed to help improve the image quality in games evenwithout the notorious full-scene anti-aliasing. If you remember, the key peculiarity of GeForce 256 GPU is the integratedgeometric coprocessor. So, NVIDIA hopes to take advantage of the chip's ability to process a great number of polygons (aswe have already mentioned the geometric engine of GeForce 256 can process 15mln triangles per second). As is known, the moretriangles are involved in the surfaces tessellation, the more natural and realistic look the objects. In other words, NVIDIAsimply suggests game developers that they should use more complicated geometric models and involve more polygons into eachscene, so that the general image quality could become better. And at the same time this measure should very beautifullysolve the problem of performance drop caused by full-scene anti-aliasing. In fact, NVIDIA suggests using over 60,000triangles in a single scene, which should increase the image quality and maintain the fps at the desired level of 60frames per second. Looks pretty nice, but no one can say for sure if the games developers will follow NVIDIA'srecommendations.

There is one more issue, which has to do with the performance: the data precision rate during scene rendering.The data is usually presented with 16- or 32-bit precision. The fps rate drops by about 10% when we shift from 16to 32-bit color depth while rendering. What causes such tangible performance reduction? The matter is that as soonas you shift to 32-bit precision rate, the amount of data you have to store and transfer gets much greater. However,the internal memory bus bandwidth is as a rule not enough, which results into noticeable delays. Besides, you shouldalso bear in mind that the time required for Z-buffer and frame buffer clearing is increasing as well. Of course, itis the problem of far not all games, it depends on the game developers, but still. As a result we see the inevitabledecrease of the overall performance. Moreover, there is one more opinion that the shift to 32-bit data precisionincreases the graphics processor utilization. As for us, we consider it to be absolutely wrong. The graphicsprocessor keeps working with the same data set at 32-bit precision as it did at 16-bit one. Both regimes are nominalthat's why we doubt if this thing can be the GPU's weak point.

And now back to the specs of GeForce 256. This time we will dwell on the GPU's 256-bit memory bus. The memory busbandwidth is twice as big as by TNT2. That is why we may suppose that the shift from 16 to 32-bit data precision willhardly make the cards based on GeForce 256 GPU suffer any performance drops. In other words, the performance is expectedto remain on the same level. But, it is only a bare supposition. Besides, we would like to point out that 256-bit memorybus is not the only thing GeForce 256 can boast: it also supports DDR SDRAM/DDR SGRAM interface. By the way, we havealready predicted that the newest graphics chips will support this graphics memory type. Our predictions seem to cometrue. You may be aware that this memory type support is not something extraordinary in the graphics accelerators massmarket. In September 1998 Rage128VR from ATI was the first graphics chip, which specs included DDR SDRAM memoryinterface support. However, memory interface support appeared to be an achievement without the proper continuation,because DDR SDRAM memory was manufactured in very small amounts and is at the same time very expensive. But DDR SDRAMcan boast a really cool advantage - twice as big bandwidth compared to the standard SDRAM memory (sometimes called SDRSDRAM). Big bandwidth is achieved due to the data transferred by both signal edges: ascending and descending. We haveevery reason to state that NVIDIA spent really much time on solving the problem of most graphics systems weakest place -of the graphics memory bandwidth. If the shift from 16 to 32-bit data precision leads to serious performance reduction inthe systems with the standard SDRAM graphics memory, the performance drop is very likely to be just minimal or even equalto zero in case of DDR SDRAM. The only disadvantage of this thing is that nobody can tell the exact day we will see thegraphics cards with DDR SDRAM local graphics memory. But if it were not for hope, the heart would break.

GeForce 256 supports AGP 4x Fast Writes, which should also positively tell on the performance in some cases. In fact,Fast Writes mode belongs to AGP 2.0 specification. It speeds up all data transfers from the CPU to the graphics processor,including:

Today's games day by day require more and more polygons. The application sends this info into the graphics processor viaAPI. The CPU controls the whole process, of course. And then the graphics processor carries out all the operations, such ascoordinates transformation, location, texturing, shading, etc. In an ordinary situation (in the systems with AGP 1.0) all thedata is transferred from the CPU through the system memory bus. The amount of data being transferred can reach up to900MB/sec! But the system bus is intended for other data streams as well, so... This problem is undoubtedly a weak pointof the system, which makes the CPU wait for the first data set to get into the graphics processor and lose its precioustime. Fast Writes mode allows the system CPU to transfer the data directly into the data graphics bus avoiding the systemmemory bus. It helps to free the PC system memory from the necessity to work with the data intended solely for the graphicsprocessor. Besides, we would like to stress that the support of this mode is not a unique occasion in the universe,which belongs to NVIDIA and to no one else. AGP 2.0 specification is open and can be used by all graphics processorsdevelopers. The secret lies in the drivers realization. And of course, you need AGP 4x support in the mainboard chipsetas well. Anyway, we will have a brilliant chance to check this mode in real life, for sure.

Another couple of comments on the graphics quality. NVIDIA suggests that the developers should calculate thereflections in real time, which can be done with cube environment mapping. This feature is supported by DirectX 7and OpenGL 1.2, and NVIDIA believes it to be an easy step towards more realistic graphics models. In generalenvironment maps are used to calculate mirror reflections and reflected lights. For a considerable while planarand spherical environment maps were the only ones used here. However, they proved unable to provide realisticmirror reflections and reflected light. Besides, the spherical maps make the developers spend too much effortand are very difficult to superpose in real time. So, the new cube environment mapping is expected to kill twobirds with one stone: it will save the developers time and trouble and will provide super real time realism.

In case of cube environment mapping, there are six planar environment maps, which form a cube around the object.The object's reflections should appear on the surface of this cube. This is where the name of the feature comes from.All this is done in order to calculate the reflections from six different directions. Besides, the reflections aresupposed to be dynamic, i.e. the considered object as well as its surrounding can be moving separately or simultaneously.As a result the environment maps can also be static or dynamic, i.e. they can be created in advance or on the fly whilethe game action is going on. As soon as the environment map is defined, it comes to textures coordinates calculation. Ithelps to find those texels, which should be superposed over some particular pixels forming the object. This object canbe of any shape and in every vertex of the triangles it consists of there is a color, which appears after theenvironment map texture is superposed.

So, cube environment mapping allows creating highly realistic reflections, which will look natural from all sides andin all directions, because the secondary lighting is also taken into account. It was impossible with the sphericalenvironment map, which considered only single direction reflection, because the map was put as a sphere around theobject. To put it short: very soon we will get much more realistic images in games. Have patience!

However, a very reasonable question crops up: does GeForce 256 have sufficient computing capacities and fillrate tomake use of cube environment mapping technique? Now we unfortunately don't have enough info about it at our disposaland all we can do is just suppose something. The only thing we can state for sure is that the environment texturescoordinates are automatically calculated without reducing the performance, while the available fillrate may be notenough at higher resolutions. Anyway, we have already devoted enough time to fillrate in the passages above. Thesituation is most likely to be OK at 1024x768, otherwise NVIDIA wouldn't have stressed that GeForce 256 GPU possessescube environment mapping hardware support.

Well, there is not so much left to talk about. One of the most interesting questions among those remaining is theDay when our curiosity and patience is finally rewarded with new graphics cards based on GeForce 256. And, you know,we are almost there already. Probably in the beginning of October, the cards from Creative, Asus, Guillemot andLeadtek and a bit later from Elsa will start conquering the market.

The first graphics cards based on GeForce 256 will have 32MB of local graphics memory, and later we can expect64MB versions as well. As for the models with 128MB of local graphics memory, nobody has made any announcements.Among the options offered will be TV In/Out, DFP interface and stereo glasses. For now we can say for sure that noone has mentioned the possible memory expansion. That's why if you have always dreamt about a 64MB version (especiallytaking into account the opinion that the cards with the memory of this size will work in the best of their ability) you'dbetter wait until it comes out. It should be worth waiting ;-) . Of course, it's pretty pleasing that the time gap betweenthe official announcement of GeForce 256 and the launching of graphics cards based on it will be very short - slightly overa month. As for the prices, a graphics card based on this chip from NVIDIA will cost $300 at the most. Is it too muchor too little? Well, let's put it this way: we are talking about a new product with a number of unique exclusive features,which suffers no competition. In this situation the price is OK. If you refer yourself to frugal users, you will probablywait for another couple of months and buy the card for only $200 or even less. Anyway, it's completely up to you. Make upyour mind!

Summing up

NVIDIA company is known as the leader of 3D graphics processors mass market. GeForce 256 GPU will undoubtedlystrengthen NVIDIA's positions in the market. Besides, this product is a serious evolutionary step forward towardsmore realistic 3D gaming graphics. This mission of the GPU became possible due to the integrated geometric coprocessorresponsible for transformation and lighting. Despite the opinion expressed by some analysts that 120MHz core frequency ofGeForce 256 was too low for the present day conditions, four-pipeline architecture easily made up for the possibleinconveniences and provided high overall performance. The available fillrate of 480mln pixels per second and 240mlnpixels per second with the enabled multitexturing will be more than enough for work at 1024x768 with 60-70fps in theexisting applications as well as in those being released during the coming half a year. The majority of users have15"-17" monitors, which automatically implies that most of them work at 1024x768. Nevertheless, NVIDIA really hopes - tosay the least of it - that game developers will use more polygons for their 3D scenes and hence 3D graphics becomesmore realistic and natural.

Besides, there is every reason to suppose that NVIDIA won't stop here. Next year, in spring, it will probably announceits new graphics processor. However, it won't necessarily be something out of the ordinary. To tell the truth, it will bemore than enough if NVIDIA finally shifts to 0.18 micron technology, which will allow increasing the core frequency of ourGeForce 256 by 30-50MHz or even more and hence the fillrate will also get higher. So, if you can summon up your patienceand wait till next spring, you won't regret it. By that time you will be not only a happy owner of a much cooler device,but you will also be able to enjoy new games using all the power of the graphics accelerator integrated into GeForce 256.

Since we started talking about GeForce's features and their support in applications, we suggest clearing out thesituation a bit. Some of the games already available in the market are now ready to make use of a number of GeForce'sfeatures. Every game with hardware rasterization, which uses API OpenGL, will be able to take advantage of the chip'scomputing capacity in terms of coordinates transformation. For example, Quake3Arena from iD Software, Inc. will feelabsolutely at home if run on a system with a graphics card based on GeForce 256. Moreover, iD is said to be modifyingits Q3A according to NVIDIA, namely it will allegedly use more polygons in the scenes. Of course, the more polygons areinvolved, the better is the image quality of the objects. Well, it seems that everything is pretty clear with OpenGL games.The only thing, which we almost forgot to mention, is the use of OpenGL Display Lists, which serve to transform the geometricdata into the format of the geometric accelerator.

And now a few words about Direct3D games. API DirectX 7.0, which is expected to be released any minute, supports geometricaccelerators, that's why any application using Direct3D coordinates transformation pipeline can be ported, so that theseoperations are carried out by T&L engine of GeForce 256. Though it will be necessary to make some changes to the applicationcode, this work will be really worth it. Some developers have already expressed their intention to introduce therequired modifications. If the application utilizes its own software coordinates transformation engine, it willalso need to be modified. It will take more time and require more effort, but only a few games need to undergo suchgrave changes. By the end of 1999 about one third of the games planned to be launched are expected to support graphicsprocessors with geometric accelerators, which will undertake coordinates transformation.

What does this all mean for the end-user? And it means that the transformation work, which belonged to your CPU'sduties will now be carried out by the geometric coprocessor integrated into GeForce 256, and as a result the systemswith a not very cool processor will be happy to get a considerable performance (fps) gain in games. We pointed out thatGeForce 256 would mostly take only transformation not for nothing. The thing is that most of the existing and ready to belaunched Direct3D and OpenGL games have software lighting, which adds to the CPU utilization instead of making the geometriccoprocessor busy. If the developers shift to hardware lighting, which is actively advocated by NVIDIA, the familiargames will impress us with additional light sources and other special lighting effects, which will contribute to therealism of the action on the monitor screen. If you remember, in the beginning we mentioned that GeForce 256 cancalculate and set up to 8 light sources for the whole scene simultaneously.

You may wonder: why, the hell, did they need to hurry so much and to launch the graphics processor with the integratedgeometric accelerator, if the existing games are unable to take full advantage of its cool features and among those games,which are about to be released, there are hardly a few ones, which will be progressive enough to use GeForce 256? The answersounds very simple: the game developers will never introduce the support for imaginary hardware. That is why the first stephad to be made by the hardware developers. NVIDIA is a leader in the graphics processors market and as every leader it hasto prove its superiority from time to time. Otherwise, the competitors will be brutal. Actually, NVIDIA could have waitedtill the total shift to 0.18 micron technology and in the meanwhile it could have also developed a special cooling systemand power supply for its GPU. But not in the leading position! This excessively explains that GeForce 256 was launched nowand not later. On the other hand, the time was very cleverly chosen: the New Year is coming pretty soon together with agreat stir around computers and component parts. GeForce 256 based graphics cards won't be able to avoid noisy success,especially without any competitors threatening its positions. And after the New Year NVIDIA will have every opportunity tofinish the technology change, to launch GeForce 256 II and to show once again who the leader is. But, this is just oursupposition, though you should admit it is not groundless.

Well, all we need now is a real graphics card based on GeForce 256 to check if our ideas are correct. There is notso much time left till this happy moment. Probably the world of games will really change, i.e. will change our idea ofthis incredible world and of the hardware we need to enjoy it in full.