X-bit labs

Creative 3D Blaster Annihilator Review

Category: Video

by FastSite

[ 11/06/1999 | 12:00 AM ]

We analyzed the GeForce256 based card features, image quality and compared GeForce's performance not only to that shown by Matrox MillenniumG400 MAX and NVIDIA Riva TNT2 Ultra, but also to the performance of a new ATI Fury Pro graphics card.

Table of contents:

So, on 31 August NVIDIA promised us to change the world. Did they succeed? Partly they did. Of course, no gravechanges were introduced but in the world of graphics accelerators life was in full swing and graphics processors (GPU) pretty successfully enteredthe mass market. Although it was S3 who first announced a new chipset with the geometric coprocessor, NVIDIA managed tocollect all the laurels due to active GPU advertising campaign and promised wonders. And what have we got in this NVIDIAGeForce256? Let's take a look at the specs:

• 120MHz core working frequency
• Up to 200MHz memory bus working frequency
• Triangle processing rate supposedly equals 15mln triangles per second; sustained DMA, transform/clip/light,setup, rasterize and render rate
• 128-bit memory bus at SDR SDRAM/SGRAM, 256-bit memory bus at DDR SDRAM/SGRAM
• Up to 6.4GB/sec memory bus bandwidth (the top value achieved with DDR memory and 200MHz bus)
• The supported memory types include: SDRAM/SGRAM (with block write support) and DDR SDRAM/DDR SGRAM
• Up to 128MB local graphics memory
• 350 MHz integrated RAMDAC with gamma correction support
• Supported display resolutions up to 2048x1536 at 75Hz
• External bus interface: full AGP 4x/2x (including SBA and DME) and PCI 2.2 (including Bus Mastering).Supports AGP 4x Fast Writes, which allows direct data transfer between CPU and GPU at up to 1GB/sec avoiding thePC system memory. It should undoubtedly have positive influence on the overall performance and help to somehow reducethe system memory bus utilization.
• I2C serial bus and DDC; ACPI D0-D3 states and PCI power management; 100% PC'99 and PC'99a compliant
• 0.22 micron technology
• 23 million transistors
• DFP interface support with scaling, filtering and resolutions to 1600x1200
• TV-out with the resolution equal to 800x600

2D Acceleration Features

• 256-bit 2D graphics engine
• Support of BitBLT acceleration, rectangle and polygon fill, line draw, hardware cursor and panning/scrolling
• Graphics acceleration at 8, 16, and 32bit color depth

3D Acceleration Features

• 256-bit rendering processor
• Integrated geometric processor for Transform and Lighting (T&L)
• 4 rendering pipelines (4 pixels per time step)
• 8 source hardware lighting for the whole scene
• Geometric engine for floating point operations and polygons location
• Fillrate: 480mln pixels per second
• Texturing rate: 480mln texels per second and 240mln texels with multitexturing enabled
• 8 texture samples for each texel
• Absolute OpenGL and DX7 support - Transform and Lighting, Cube environment mapping, projective textures and texture compression
• Supports all 5 formats of DX6 texture compression
• Hardware bump mapping support. Embossing, Dot Product.
• Programmable blending of multiple textures
• Vertex Blending support, which provides natural-looking complex images in places of polygons connections. For example, if you need a model of a crankshaft taken from the car engine, vertex blending will provide natural looking fractures, otherwise all the polygon joints will be very noticeable.
• Rendering in 32bit color depth regime
• Full scene anti-aliasing
• 8-sample anisotropic filtering
• 8-bit stencil buffer support
• 16/24/32bit Z-buffer support
• Specular lighting and diffuse shading
• Enhanced alpha blending modes
• Vertex and pixel fog
• Supports the textures of 2048x2048 texels with 32bit precision

DVD/DTV Support

• HDTV (High Definition Television) support
• Video overlays support.
• Vertical and horizontal scaling with the coefficient equal to 8.1
• Hardware brightness, saturation and color contrast controls
• Multiple HDTV configurations supported via DMA host port and bus mastering
• Motion compensation support (we hope it will work)
• Hardware subpicture blending and highlights
• Hardware color spaces transformation support (YUV 4:2:2 and 4:2:0)
• Multi-pass filtering of video images (5 horizontal and 3 vertical levels)
• Data transformation from the planar YUV12 (4:2:0) format into the packed 4:2:2 format and vice versa

As we can see, the features list is so rich that we start worrying about the proper realization of all the announcedfeatures. First versions of Savage2000 launched by S3, for instance, appeared with a bit reduced frequencies range comparedto what we saw in the initial specs (they explain it with low percentage of fit chips).

We managed to find one of the first graphics cards on NVIDIA GeForce 256 chip - Creative 3D Blaster Annihilator. This cardlooks very much like its predecessor Creative 3D Blaster Riva TNT2 Ultra, of course has an AGP form-factor, AGP 2x/4x interfaceand the same 32MB 5.5ns SDRAM located in 16 microchips on both sides of the card:

So, what is the most striking thing? It is the memory similar to the way it was made on Creative 3D Blaster Riva TNT2Ultra - all chips are the same, we even compared the marking. As we have already mentioned the bottleneck of most today'sgraphics accelerators is the memory bus bandwidth. In other words, no matter how big the bottle is: if the neck is the samesize as by a much smaller one, the amount of water, which can pour out of it in a certain period of time remains the same inboth cases. This is the first cause for disappointment. What is the sense of making a more powerful chipset if the memory failsto catch up with it anyway? The sad fate of 32-bit color remained sad again. We have every reason to say so since at this colordepth the data exchange rate with the memory is too low for the graphics accelerators to perform up to the required level.

Besides, we would also like to note that GeForce 256 is provided with 23 million transistors and manufactured by 0.22micron technology. It looks as if NVIDIA simply hasn't had enough time to finish the transition to 0.18 micron technology.We think that a large number of transistors used and 0.22 micron technology was the main reason for NVIDIA to restrictGeForce's core frequency to 120MHz because otherwise there could occur some problems with chip overheating and power supply.In fact, NVIDIA GeForce 256 GPU based graphics card really gets much more overheated without any additional externalcooling than its fellow based on NVIDIA Riva TNT2 Ultra chip.

Of course the most remarkable innovation of NVIDIA GeForce 256 is the integrated geometric coprocessor. You can arguethat the best place for it is inside the chip and not outside, as an external microchip, - each solution has its positiveand negative sides. The primary advantage of the chosen location is the cheaper design of a reference graphics card andquicker chipset launching into the market. The main drawback is the limited clock frequency for instance, caused by the chipoverheating (coprocessor makes a significant contribution here).

In 3D rendering, after all the required data is already obtained, it goes to coordinates transformation stage where theinitial coordinates of 3D-objects get transformed into 2D-ones used for the scene we see on the monitor. It is important toremember that each point can be lit by several lights. And all this information should be applied with regard tothe viewing angle, the so-called correct watching point and perspective. Only after that the final result is displayed on themonitor and at the same time all the invisible details of the scene get deleted (clipping), so that to save the furtherrendering time and effort. Transformation and lighting operation require very large computing capacities. These operationscan be carried out either by a CPU or by a special geometric coprocessor. And if the latter is available, then the uninvolvedresources can serve other gaming purposes, such as the realization of more progressive AI or environmental physics of the modeledworld (gravity force, air density, etc.).

The geometric part of GeForce 256 responsible for T&L and Clipping boasts really impressive abilities. Just imagine:GeForce 256 can provide 15 million textured polygons with Z-buffer every second! It's an enormous amount. At the sametime the system CPU can be absolutely free from transformation and lighting operations, which usually take about 70-90% of theCPU resources. In this respect an urgent question arises: is there any serious need in graphics cards with the GPU rightnow? As to the applications developers, in about a year there will hardly be any games left working without a graphics cardwith an on-board geometric accelerator. They state that the difference between the performance of the graphics acceleratorwith the geometric coprocessor and the one without it is just incredible, which leads to considerable difference in imagequality. As for the practical importance of hardware vertex coordinates transformation with the help of a geometric coprocessor,the thing seems more or less clear. All OpenGL and Direct3D games (slightly modified) can make real use of this feature. However,as it comes to lighting, some problems still crop up.

The first lighting source is calculated by the geometric coprocessor integrated into GeForce 256 "for free", i.e. withoutany performance drop, while all the rest (which is 8 light sources altogether) theoretically reduce the overall performancethough not too much: by 5-7% at the most (we will return to real state of things later in our review). The game developersconsider vertex lighting to be carried out too stencil-like. But the technologies and tools used for that keep developing,and besides according to NVIDIA there is still enough time since they predict vertex lighting to be widely spread only by theend of next year.

Then, GeForce 256 can boast 4-pipeline architecture (wanted just to remind you that NVIDIA Riva TNT/TNT2 have only 2 ofthem). It means that every time step there can appear 4 pixels simultaneously. Since the GPU clock frequency is 120MHz, thefillrate will make 480 million pixels per second. However, the pipelines of GeForce 256 have one texturing block each, whichmeans that at the same clock frequency equal to 120MHz the fillrate will drop down to 240 million pixels in applications usingmultitexturing (when two textures are simultaneously superposed over one pixel). Note that only filtered textures aresuperposed 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 interesting solution, however, the samples seem to beinsufficient. It could be much better to involve 16 or even 32 texture samples to define the final texture color. Inthis case the final color definition requires 8 samples, which should in the end provide high quality image. Why doesNVIDIA use only 8 samples and not more? Because this realization can be carried out with one texturing block only, whichwas made in GeForce 256. However, even with only 8 samples the image quality gets much better. We even dare suppose thatthe image quality of the existing games can also improve. This statement is illustrated by Quake3 screenshots (for a bettercomparison we indicated all MIP-levels with different colors). In case you forgot: MIP-mapping serves to make the objectdetailization depend on the distance between the object and the observer (the farther is the object the fewer details wesee and the more washed out are its contours). That is why the whole scene is divided into MIP-levels, which borderingedges should be smoothed otherwise the whole 3D-scene will lack realism. There is number of filtering techniques, which are utilizedfor this particular purpose. We are going to devote them a bit more time in our review.

NVIDIA GeForce 256

NVIDIA Riva TNT2

Take a look at the difference between the filtering quality in both cases. The screenshots below represent the samepicture but this time without any filtering so that you could clearly see the MIP-levels:

NVIDIA GeForce 256

NVIDIA Riva TNT2

Have you noticed the shape of those MIP-levels? Those of NVIDIA GeForce 256 are rounded and are located as a semicirclewith a player in the center. Of course, this scene division into MIP-levels seems to be the most natural one because theterritory within human sight also forms a semicircle. And by NVIDIA Riva TNT2 these levels are somewhat angular. Note thatthese results were obtained with the same drivers and hence the problem was solved in hardware, which is really pleasing.

And now let's say a few words about fillrate. This factor directly influences the performance of any graphics adapter.So, after GeForce 256 announcement a lot of editions expressed their concern that low graphics core frequency would lead tolow gaming performance. A bit later NVIDIA's competitors represented by 3dfx CTO Scott Sellers also shared their opinion onthe matter. In particular, Scott said that GeForce's fillrate was too low and insufficient for the card to be used in today'sgames at high resolutions. Of course, he didn't forget to mention that the future 3dfx product would boast a much higherfillrate and would allow playing games at high resolutions without any performance losses. In fact, according to the numerousrumors about GeForce 256, its fillrate was allegedly promised to be 3-4 times higher than that of TNT2. However, it neverhappened. For instance, some cards based on TNT2 Ultra can boast the fillrate of 190mln pixels per second with two texelssuperposed over one pixel (in games with multitexturing). As for GeForce 256, in the same games its potential fillrate is240mln pixels per second with two texels superposed over a pixel per time step. You see that the gain is quite insignificanteven theoretically. Nevertheless, it is still hard to answer the question: "What is better for a graphics chipset: highfillrate or a geometric coprocessor?" The coprocessor is for sure a very attractive feature of the future graphics cardssince it is taking over all routine calculations. It is almost impossible to redesign the existing games (with the onlyexception of a few OpenGL games) for the new GPU. And with a higher fillrate it appears possible to sharply increasethe fps rate in the today's games and to let the users play them at higher resolutions. Take for example two well-knowncompanies: Epic (Unreal Tournament developers) and id Software (Quake3 developers) who have absolutely contrary opinionson this problem. In the first case the stress is made on the performance, i.e. a more powerful accelerator is needed, whilein the second case - since the GPU carries out the necessary calculations - it is the higher scene complexity that becomesof paramount importance and hence requires an accelerator equipped with a GPU. Undoubtedly, we are talking about the futureprojects, we don't mean the today's games. Although it is very illustrative that id Software decided to make its Quake3utilize the graphics cards geometric coprocessor to as full extent as possible, and Epic refused point-blank to introduceany changes into Unreal Tournament concerning its optimization for GPU (probably it is Epic's love to 3dfx that tells :-)).Here is a fragment from Quake3 where such commands as r_subdivisions 1 and r_lodCurveError 10000 help to get smoother surfacecurves dividing them into a larger number of polygons while the overall performance of the card hardly suffers:

The increased amount of polygons

The default amount of polygons

You won't deny that the card's performance also depends on such factors as additional special graphics effects.In particular, full-scene anti-aliasing is expected to cause fillrate drop and hence lower fps in games. It is verylikely to be so, actually. But NVIDIA stakes another GeForce's feature, which is supposed to help improve the imagequality in games even without the notorious full-scene anti-aliasing. NVIDIA tries to take advantage of the chip'sability to process a larger number of polygons. As is known, the more triangles are involved in the surfaces tessellation,the more natural and realistic look the graphics objects. In other words, NVIDIA simply suggests game developers that theyshould use more complicated geometric models and involve more polygons into each scene, so that the general image quality couldbecome better. And at the same time this measure should very beautifully solve the problem of performance drop caused byfull-scene anti-aliasing. In fact, NVIDIA suggests using over 60,000 triangles in a single scene, which should increase theimage quality and maintain the fps at the desired level of 60 frames per second. Looks pretty nice, but no one can say forsure if the game developers will follow NVIDIA's recommendations. Nevertheless, we suggest looking at the screenshot belowtaken from one of NVIDIA's demo programs:

Here you can clearly see that it is now possible to create an almost fully realistic model of a human face, which shallundoubtedly raise the mark for such games at 3D-shooters and the like. Of course, we can say that the top perfection isreached if the modeled faces acquire realistic mimics - another valuable contribution to higher naturalism of the gameslike that.

And now we would like to dwell on the water :-) Yeah, this object of 3D-world is of great interest to lots of designersfirst of all due to its highly dynamic character. Unlike the walls for instance, water is constantly changing. Matroxsuggested using EMBM (Environmental Mapped Bump Mapping) to make the water look realistic. We don't deny the uniqueadvantages of this method and agree that it provides the required effect. However, unfortunately, only Matrox G400chipsets possess this technology that is why most game developers take their time and don't strive for immediaterealization of this technology in their products. And is there any good way of improving water realism without anysuper technologies? Yes, there is. However, in this case the processor needs to have really high computing capacity.And it is exactly here that the geometric coprocessor comes to the rescue. The three screenshots below show the qualityof the modeled water surface with different reflective and semitransparent effects enabled:

What do we see? Dead calm. And what happens if we slightly disturb the water?...

Well, the waves did a good deal of work and contributed to the scene realism. But how? Well, this amazing effect wasachieved due to the displacement map looking like a sort of grid with the knots denoting the environment transparency values.Light reflection and refraction vectors were also obtained. Thus, with relatively powerful computing resources of the graphicscard this method turns out available not only to chipsets supporting unique technologies but also to all the rest as well, andhence it can be used for 3D-games development.

We continue. NVIDIA suggests that the developers should calculate the reflections in real time, which can be done with cubeenvironment mapping. This feature is supported by DirectX 7 and OpenGL 1.2, and NVIDIA believes it to be an easy step towardsmore realistic graphics models. In general environment maps are used to calculate mirror reflections and reflected lights. Fora considerable while planar and spherical environment maps were the only ones used here. However, they proved unable to providerealistic mirror reflections and reflected light. Besides, the spherical maps made the developers spend too much effort and werevery difficult to superpose in real time. So, the new cube environment mapping is expected to kill two birds with one stone: itwill 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'sreflections should appear on the surface of this cube. This is where the name of the feature comes from. All this is done in orderto calculate the reflections from six different directions. Besides, the reflections are supposed to be dynamic, i.e. theconsidered object as well as its surrounding can be moving separately or simultaneously. As a result the environment mapscan also be static or dynamic, i.e. they can be created in advance or on the fly while the game action is going on. As soonas the environment map is defined, it comes to textures coordinates calculation. It helps to find those texels, which shouldbe superposed over some particular pixels forming the object. This object can be of any shape and in every vertex of thetriangles it consists of there is a color, which appears after the environment map texture is superposed. So, cube environmentmapping allows creating highly realistic reflections, which will look natural from all sides and in all directions, because thesecondary lighting is also taken into account. It was impossible with the spherical environment map, which considered only singledirection reflection, because the map was put as a sphere around the object. To put it short: very soon we will get much morerealistic images in games.

Here is what we see according to one of the demo-programs provided by NVIDIA:

Spherical map

Cubic map

The difference is more than evident.

And now a few words about the card itself. The installation of Creative 3D Blaster Annihilator ran without any problems,the drivers were easy to install as well. By the way, starting from 3.* version NVIDIA changed the drivers design and contents.That's why let's talk a bit more about new drivers. We want to draw your attention to the fact that NVIDIA seems to be continuingits experiments about removing important options from the drivers, which were first started in drivers version 2.*. If you installdrivers ver. 3.53 there will be no way for you to manipulate Vsync regime (monitor refresh rate synchronization with RAMDAC).Besides, the old solutions successfully used for Registry editing in 2.* versions don't work any more. That's why we made upour mind to offer you the key you have to include into Registry in order to enable all the options unavailable as default:

REGEDIT4 [HKEY_LOCAL_MACHINE\Software\NVIDIA Corporation\Global\NVTweak] "CoolBits"=dword:00000003 

As soon as you add this key, you get the opportunity of Vsync setting:

And you also get a page for memory working frequencies setting (overclocking):

Take a look at the remaining pages:

This panel displays brief graphics card info and drivers components versions:

This page is devoted to Direct3D settings. Table fog and z-buffer emulations are enabled as default. It is very likely thatthe laurels of 3dfx shattered NVIDIA's peace and quiet with their logo was displayed during Glide initialization. In thisrespect NVIDIA decided to add an option displaying its own logotype while Direct3D programs were running. And as for the other driverssettings, they hardly differ from 2.* version except the absence of tri-linear filtering replaced by 8-pixel anisotropic filteringin MIP-mapping section.

This page deals with OpenGL settings. There is the whole bunch of different opportunities for filtering setting, color depthchoice (before that 16- or 32-bit color used to be closely connected with the desktop color depth). Besides, the option for Vsyncenabling/disabling for OpenGL applications is also worth mentioning.

Well, here we would like to finish with the peculiarities of NVIDIA GeForce 256 chipset and the first graphics card based on it.Note that Creative card was initially intended as a reference design, i.e. as an example recommended for further cloning by othermanufacturers.

And finally the tests run on the following testing system:

• Intel Pentium III 600 CPU;
• ASUS P3B-F mainboard;
• 128MB PC100 system memory;
• Quantum FB CR 6.4GB HDD;
• ViewSonic P810 monitor;
• Windows 98.

For a few tests we also used another PC configured as follows:

• AMD K6-2 450 CPU;
• Chaintech 5AGM2 mainboard;
• 128MB PC100 system memory;
• Quantum FB CR 6.4GB HDD;
• Nokia 447Xav monitor;
• Windows 98.

The tests were carried out with the driver version 3.53. And the results shown by our hero - Creative 3D Blaster Annihilator -were compared to those of its main competitors: Matrox Millennium G400 MAX (driver version 5.30), Hercules Dynamite TNT2 Ultra(NVIDIA Riva TNT2 Ultra, driver version 3.53) and an absolutely new card for us ATI RAGE FURY PRO (ATI RAGE 128 PRO, driver version6.30CDH34e).

Let's start with 2D graphics. Speaking about the image quality here we can't help mentioning high RAMDAC frequency, whichprovides brilliant image quality at resolutions including 1600x1200. Unfortunately, we didn't have the chance to try higherresolutions as well, because there was no monitor in our lab supporting such regimes. The performance was taken at 1600x1200 in 32-bitcolor with the help of WinBench99. The results are given in the following table:

As we can judge by the performance, our graphics card left all the latest novelties behind. Even Matrox MillenniumG400 MAX failed to retain its positions. That's why we have every right to state that Creative 3D Blaster Annihilatorhas no match in terms of universality and high performance in 2D. Here we mean only the leadership among gaming graphicscards - an important remark for professionals working with the finest 2D elements. However, as for us, our eyes provedunable to detect any differences between the image quality shown by Creative 3D Blaster Annihilator and the alreadymentioned Matrox Millennium G400 MAX.

And now to 3D-graphics, and in the first place we would like to tackle the performance. This time we veered away fromour traditional plan and added a number of other items. To begin with, take a look at the results obtained in NVIDIATreeMark - a benchmark symbolizing NVIDIA GeForce 256. This demo-test was initially intended for this particular chipset and thewhole set of its features. It presents the graphics accelerator in really tough conditions (with numerous light sources and largeamount of polygons). This program runs under OpenGL interface:

Let's compare the performance of NVIDIA GeForce 256 and NVIDIA Riva TNT2 Ultra at different resolutions with the PentiumIII frequency set to 600MHz (the test was run with all default settings: the level of depth complexity equal to 5, the numberof light sources - 4):

It's evident that NVIDIA GeForce 256 proved simply super giant. Now we have a very clear picture of what it may be ingames with hardware T&L.

Take a look at NVIDIA GeForce 256 working at different processor frequencies (Pentium III - 600 and 450MHz, Pentium II -300MHz, K6-2 - 450MHz). The benchmark default settings are the following:

Well, the use of the geometric coprocessor proved very fruitful because as we see the results at all working frequencies(except K6-2 - 450MHz) are absolutely identical! So, we can say once again that if there were games squeezing the maximum outof these GeForce's features, we could easily achieve impressively good results even with less powerful CPUs. But why did K6-2fall so greatly behind? This question is not very simple to answer. Probably it is the imperfect AGP bus plus poor CPU's supportof those graphics card features, which have nothing to do with T&L.

Let's try to follow the dependence of NVIDIA GeForce 256 performance on the quantity of light sources while the scenecomplexity (depth = 5 as default) remains unchanged. The tests were fulfilled at 1024x768:

As we can see, the performance of NVIDIA GeForce 256 goes along the descending curve if the number of light sourcesgets bigger. It is evident that only few light sources cause a 7% drop as it was promised, while in most cases it makesover 10%. If we consider GeForce 256 in case of 8 light sources its lowest performance will be about 1.7 times smaller thanthe highest one. But if we take NVIDIA Riva TNT2 Ultra then the drop will make about 3 times. Besides, GeForce 256 still haschances for improvement with the new drivers versions.

Now let's check the dependence of NVIDIA GeForce 256 performance on the level complexity, namely on the number of polygonsin the current frame. The tests were run for cases with 4 lights (default) at the resolution set to 1024x768:

This graph very clearly shows that NVIDIA Riva TNT2 Ultra loses its fastness even if the scene complexity is just slightlyincreased, while in case of NVIDIA GeForce 256 the situation looks not so tragic although the curve sharply goes down. However,we have to bear in mind that one of the possible reasons for such considerable performance drop can also be a very narrow memorybus.

For our next test we used another novelty - Unreal Tournament (3.48 demo-version) from Epic Games. Unfortunately, this gamedoesn't have any official benchmarks that is why we simply saved a special file mydemo1.dem, which you can undoubtedly try athome. We would like to point out right away that this demo-thing wasn't created for bot games. Its main goal was to test thegraphics cards that is why you may notice some really sharp and awkward movements, shooting without any definite target. Weplayed the game with "God" so that to avoid static screens (we mention all these things in order to prevent the inveterategamers from expressing their indignation at the quality of our game :-)) The gaming episode is pretty long. It lasts about7-10 minutes and finishes with a total victory on one of the levels. It allows us to be more objective when obtaining thefinal value characterizing the cards' performance.

Have a look at the figures. No doubt, NVIDIA GeForce 256 won the first prize for the highest performance, however, thedifference is very insignificant. This game is known to use no hardware T&L and the performance depends only on the fillrateprovided by the graphics card used. But even if the performance is much higher, NVIDIA GeForce 256 fails to show the appropriategain. Why so? Because of the CPU. Yes, Unreal Tournament is a game of the future: even a 600MHz Intel Pentium III is not enoughfor it.

And here we got something new and unexpected: the graphics card based on NVIDIA GeForce 256 appeared behind that based onNVIDIA Riva TNT2 Ultra. No matter that the difference makes just a few fps: it is still pretty meaningful. The determinativehere is very likely to be the graphics memory working frequency, which plays a very important role for the performance in32-bit color - 200MHz by NVIDIA Riva TNT2 Ultra against 166MHz by the graphics card based on NVIDIA GeForce 256.

Here we feel like veering slightly away from the main topic of our discussion and saying a few words about overclocking.As a rule, we test all graphics cards not only at their nominal frequencies but also when overclocked, in order to make thepicture as full as possible. Unfortunately, NVIDIA GeForce 256 chip is extremely subject to overheating that is why it can'tboast impressive overclockability. We managed to make the graphics card work rather stable only at 135/185MHz (chipset andmemory frequency correspondingly) and only with additional external cooling.

And now we will offer you the results achieved by the graphics cards in the traditional testing tools. These games are:

• Rage Expendable;
• Monolith Shogo (demo - revshogo);
• id Software Quake2 v.3.20 (demo - massive1.dm2);
• id Software Quake3 Test v.1.08 (demo - q3demo1.dm3).

Expendable shows our graphics accelerators working in Direct3D with multitexturing disabled. We utilized this tool to carryout four tests.

You have probably noticed that NVIDIA GeForce 256 succeeds in surpassing its rivals only at higher resolutions.It looks as if all the resprictions were imposed by the computing capacity of the system CPU even in case we have aPentium III 600MHz system. Besides, another new graphics card left our hero somewhat behind at a number of resolutions.This newly born leader was ATI RAGE FURY PRO.

Here the picture strikes as really shameful for the graphics card based on NVIDIA GeForce 256. The card on ATI RAGE 128 PROproves its indisputable leadership, while the card on NVIDIA Riva TNT2 Ultra evidently performs on the level of NVIDIA GeForce256. We can't think of any reasonable explanation for this phenomenon. To tell the truth, two ideas occurred to us: the firstabout a slower memory than that of NVIDIA Riva TNT2 Ultra and the second about a cleverer optimization of ATI RAGE 128 PRO for32-bit color regime. Anyway, we think that ATI Company deserves our most sincere congratulations. The new graphics card fromATI managed to leave behind not only all former graphics leaders but also a very promising graphics card based on a fresh newchipset GeForce 256 with the highest fillrate of all. ATI did a good job!

Now let's discuss the performance drop caused by the shift to 32-bit color depth:

Frankly speaking, the previous test gave us every reason to expect these results. ATI RAGE 128 PRO made our last doubtsabout its leadership vanish having shown the best results in 32-bit color. Then we can welcome Matrox Millennium G400 MAX andthen - strange as it might seem - comes the card on NVIDIA Riva TNT2 Ultra. As far as the card based on NVIDIA GeForce 256 isconcerned, it occupies the last place in this list. Just take a look at the performance drop. Well, it is another evidence thatour suppositions mentioned in the previous passage and in the very beginning of the review were not ungrounded. And namely weexpressed concern about such a narrow memory bus combined with such a powerful chipset. It's simply blasphemy! A reproachaddressed directly to NVIDIA Company. Two-pipeline card equipped with a slower memory in terms of working frequency (ATI RAGE128 PRO works at 140MHz chipset and 159MHz memory frequency) managed to prove cooler than a four-pipeline graphics card!Unbelievable! And even if we take into account that the multitexturing is enabled, which means that in fact there are twiceas fewer pipelines involved, this is still a very poor excuse for NVIDIA GeForce 256. It can't make up for its failure.

It was on purpose that we provided this diagram together with the similar results shown in TreeMark, which makes use ofthe GPU. We wanted to demonstrate that if the CPU frequency decreases in the games not requiring cool features provided bythe geometric coprocessor, the performance of the graphics card based on NVIDIA GeForce 256 gets lower the same way it doesby the graphics cards based on NVIDIA Riva TNT2 Ultra.

Further. Shogo is a very remarkable game especially due to the possibility to enable/disable multitexturing regime that'swhy we carried out two tests:

The situation here is quite ordinary: the graphics card on NVIDIA GeForce 256 significantly surpasses its competitorsat higher resolutions. It's interesting and even strange that this test turned out the worst for ATI RAGE 128 PRO.

This is a very advantageous situation for NVIDIA GeForce 256 based card. This is an excellent example of good realizationof this regime in GPU GeForce 256. Note that Shogo runs in 16-bit color, so if we compare these results with the resultsachieved in Expendable in the same color regime, the similarity between them will be hard to deny. Actually, we assume thatalthough this multitexturing regime ensured some performance gain for the previous NVIDIA chipsets, their fame dwindled becausethere were some other chipsets, which could boast higher performance gain due to multitexturing. In this case NVIDIA did a finejob trying to solve this problem, so that multitexturing is no longer an Achilles' heel of GeForce 256. This fact is of greatimportance especially since the number of games supporting multitexturing keeps taking on.

And now to Quake2. We decided to start our OpenGL investigation with this game. The tests were fulfilled in 16 and 32-bitcolor.

If we take into consideration that NVIDIA has been quite successful in OpenGL recently, the brilliant results will hardlyseem surprising. The card on NVIDIA GeForce 256 left everyone far behind.

Although the graphics card based on GeForce 256 managed to strengthen its positions here the gap between it and itsrivals reduces noticeably at higher resolutions. Is it again the memory? The narrow bus bandwidth? Well, looks very muchlike that... Here we have to seek justice among you, our readers: NVIDIA undertook a very brave thing having introduced inits latest drivers compulsory disabling of multitexturing in 16-bit color and its compulsory enabling in 32-bit regime.Unfortunately, we have no chance to test NVIDIA GeForce 256 with the older drivers ver. 2.08 (they simply don't recognizethis chipset), though we can say for sure that NVIDIA Riva TNT2 Ultra performed much worse with the new drivers (ver. 3.53)than it used to with the older drivers (ver. 2.08). and this drop was quite significant in 16-bit color regime. Besides, ifthe enabled ForceMultitexture key used to be quite efficient, now it plays almost no role. That is why can't find suitablewords to describe our negative attitude towards this move. In fact we can't make out the purpose of hiding these options fromthe users. They simply considered us to be absolute idiots "unable to understand the universal truth". Why on earth did theyhide all the operations for proper work in 3D?

As for Quake3, it is now represented by a test demo version 1.08. The integrated into the game itself Q3demo1 is intendedsolely for testing purposes. There are four standard modes that's why we made four tests and one larger investigation ofGeforce's performance dependence on the CPU used in the system.

The results obtained here are very similar to those we saw in Quake2 in 16-bit color. All proportions remain and theleadership again belongs to NVIDIA GeForce 256. Overclocking made its irreplaceable contribution to the card's performance:it quickly grew up. We can't help mentioning a very weak ICD OpenGL by ATI RAGE 128 PRO. They should have felt ashamed torelease such poor software on the threshold of the XXI century! If we compare these results with those in Direct3D, thesoftware developers' fault will be more than evident. If they only created a really good ICD OpenGL...

And again NVIDIA GeForce 256 can boast nothing but the minimal performance gain compared to NVIDIA Riva TNT2 Ultra.

No comments. Everything is the same as in Normal mode.

And again nothing changed. The only thing that catches our eye is the lack of the system CPU computing capacity.

Let's try to find out how the performance of NVIDIA GeForce 256 depends on the processor frequency in vertex lighting:

As we have mentioned before, id Software is organizing its work on Quake3 so that to make the geometric coprocessorresponsible for some transformation and lighting calculations without any specific changes. In particular, it has to dowith vertex lighting. Quake3 supports two types of lighting: with the lightmaps and vertex lighting. Although the latteris pretty poorly carried out in Quake3 can still be entrusted to the GPU instead of the central processor. The chart aboveshows that in case the CPU frequency decreases the performance doesn't drop all at once as if it were a game using noadvantages provided by T&L.

So, what will our general conclusion look like according to the testing results? First of all, we have to admit thatthe card didn't show the so greatly expected revolutionary performance. Of course, we do not deny the performance gain(it would be surprising if there were no gain at all) especially taking into account the fillrate equal to 480Mpixels persecond. (If you remember NVIDIA Riva TNT2 Ultra in Hercules Dynamite TNT2 Ultra has only 350Mpixels/sec.) However, as wesaw NVIDIA GeForce 256 based graphics card performed not very well and sometimes failed to surpass even its predecessor.And as for 32-bit color, the advantages there are even fewer. We were very disappointed with the sad fact that a really coolchipset suffered a total failure because of the memory bus small bandwidth. Nevertheless, we noticed that in games supportinghardware T&L this chipset proves more and more powerful. Since a lot of games like that are about to come out pretty soon webelieve that graphics cards based on NVIDIA GeForce 256 will have a very happy and successful future even with the today'smemory bus bandwidth. And if the next graphics cards on this chipset use DDR DRAM memory their future seems to turn out evenmore promising. Unfortunately, we can't say anything else on the matter yet, but as soon as we get a graphics card with DDRSGRAM on board, we will definitely share our opinion with you.

Speaking about the quality of 3D graphics GeForce appeared up to our expectations. The image quality was similar to thatwe got with NVIDIA Riva TNT2 Ultra, i.e. it was pretty high. Besides, the developers finally removed the fly from the ointment,which spoilt the whole impression about the graphics quality provided by NVIDIA chipsets. This irritating fly was tri-linearapproximation. When we discussed the peculiarities of NVIDIA GeForce 256 chipset we offered you a couple of Quake3 screenshotswith a perfect anisotropic filtering that's why we can be absolutely sure that this chip will never upset us with these "stupidpoints" falling out of the general picture. And the best way to estimate the overall quality is to look once again at thescreenshots above.

Passing over to the final part of our review we would like to point out one more peculiarity of NVIDIA chipset. It isDVD-Video. As we have already mentioned in some of our reviews devoted to the graphics cards on NVIDIA Riva TNT2, thesechipsets hardly provided any real DVD-Video support. The CPU was always so greatly utilized! Although the images were veryclear-cut (they used the system of overlays).

Not so long ago a new software player WinDVD ver. 1.2.85 came out. With its help we managed to get the CPU utilization ofless than 45% when playing a DVD-disk on NVIDIA GeForce 256 based graphics card! It was unbelievable! And it was really muchbetter than the results obtained on other graphics cards on NVIDIA chipsets. Moreover, the image quality didn't suffer at all! Weconsider these results to be a really important achievement, which makes the graphics cards based on NVIDIA GeForce 256 evenmore universal.

Summing up.

The new graphics chipset NVIDIA GeForce 256, which symbolizes a new era of 3D computer games, stood the first tests. Thischipset contains huge potential and allows redirecting some routine calculations to the graphics coprocessor, which willundoubtedly let game developers focus on higher quality of gaming elements, introduce new effects (such as Cube environmentmapping), increase the number of polygons used for scene striving for higher realism and more natural look of the createdmodels. We can't praise NVIDIA GeForce 256 for the way it performed in the today's applications, especially in 32-bit colordepth. Besides, NVIDIA Company proved very inconsistent in providing their drivers with the necessary options. However, youshouldn't miss out the fact that this new product costs much cheaper than the first graphics cards on NVIDIA Riva TNT2 Ultra.This advantage undoubtedly makes NVIDIA GeForce 256 GPU and the corresponding graphics cards a really attractive choice. Ofcourse, we understand very well that the today's market is simply buried under the whole bunch of all sorts of graphics cardsand it may appear very difficult to comfortably arrange a new transitional product there. You are probably wondering why wecall it a "transitional" product, aren't you? In fact we are sure that in spring of the year 2000 the competition will aggravateeven more. And tis upsurge won't be caused only by a new product from 3dfx, but also by a new improved NVIDIA GeForce 256 and some mysteriousthing from Matrox. Well, let's have patience. Time will show everything. And in the meanwhile if you are longing for a graphicscard on NVIDIA Riva TNT2 Ultra then why don't you change your mind and decide on a new NVIDIA GeForce 256 GPU?