NVIDIA GeForce4 GPU Review

by FastSite
02/06/2002 | 12:00 AM

On February 6, 2002 NVIDIA officially announced a new GeForce4 chip also known as NV25.<%BANNER[article]%>

There will be a new High-End graphics cards family built on this chip. The family will be called NVIDIA GeForce4 Titanium. At present there are two products in this family:

• NVIDIA GeForce4 Ti4600. This graphics solution will have 128MB DDR SDRAM with 128bit width. The chip will work at 300MHz, the memory - at 650MHz.
• NVIDIA GeForce4 Ti4400. This graphics product will support 128MB DDR SDRAM with 128bit width and will work at 300MHz chip and 550MHz memory frequencies.

Probably, soon after the mentioned above models come out, we will also see GeForce4 Ti 4200 graphics cards equipped with slower graphics memory and slower GeForce4 chips working at lower frequency. So far, we have no idea what working frequencies GeForce4 Ti4200 may have, as there are too few defective chips now :)

No secret that many people pinned a lot of hopes upon the launching of GeForce4. Some of them expected nearly a revolution in 3D graphics, or at least a serious dash forward. Getting a little bit ahead of time, we have to say that many of these hopes will not come true. NVIDIA GeForce4 can't boast as revolutionary innovations, as GeForce3 a year ago could.

However, we will never deny that the developers created a new fast and dangerous 3D graphics monster having added many new features to the already common GeForce3 architecture, which has become the basis for GeForce4.

Taking all the security precautions we decided to find out now what will the new generation of these monsters bring to the world of 3D graphics and got hold of a GeForce4 Ti4600 sample for our investigation :)

Closer Look: GeForce4 Chip

Well, here we will consider the differences between GeForce4 Titanium and GeForce3 Titanium families. For your convenience we summed up all the basic features of GeForce4 Ti4600, GeForce3 Ti500 and ATI RADEON 8500 in a table:

Now let's dwell on each innovation introduced in GeForce4 and see what else besides the working frequencies distinguishes it from GeForce3.

Vertex Shaders Unit

NVIDIA GeForce4 feature 2 Vertex Shaders Units, which are completely analogous to those used in GeForce3. Vertex Shaders units of NVIDIA GeForce4 work in parallel, like those of the Microsoft X-Box graphics core. As a result, the vertex shaders processing speed increases.

Let's check if this statement is true with the help of 3DMark2001 synthetic benchmarks:

You can see the advantage of NVIDIA GeForce4 Ti4600 with a naked eye: the doubled T&L unit of the new NVIDIA chip turns out faster even if its working frequencies get reduced down to those of GeForce3: 240MHz chip and 500MHz memory. And at its nominal frequencies… no comments are necessary, we believe.

The situation with the Vertex Shaders is not so optimistic any more. Despite the remaining advantage over GeForce3 Ti500 in this test, GeForce4 Ti4600 falls behind ATI RADEON 8500. Highly efficient Vertex Shaders unit of ATI RADEON 8500 allowed it to remain the leader here and to leave behind GeForce4 Ti4600 with its two Vertex Shaders units and higher chip clock frequency.

Pixel Shaders Unit

As you can see from the table, NVIDIA GeForce4 supports Pixel Shaders version 1.2 and 1.3. In fact, even the GeForce3 core can work with Pixel Shaders version 1.2 and 1.3, however, the drivers report that GeForce3 supports only Pixel Shaders version 1.1.

GeForce4 borrowed the Pixel Shaders unit from GeForce3 without any architectural changes, which you can see from the configuration of the available registers in versions 1.1-1.3:

The number of the available registers remained unchanged. However, the maximum number of registers allowed entering the instructions as initial data has been increased. Also, there appeared some new instructions:

New instructions allow performing some "more advanced" vector and scalar calculations. But the most interesting instruction of all seems to be texm3x2depth (some kind of analogy to texdepth for Pixel Shaders version 1.4 supported by ATI RADEON 8500). Taking into account the data received at the previous stages of the Pixel Shader and the initial Z value for the pixel processed, texm3x2depth instruction will derive the new Z value for this pixel.

This tool for varying the depth value for each pixel can be used together with the bump mapping methods. In this case, the intersection line of the bumpmapped polygons will no longer be straight and will not ruin the bump mapping effects. The whole thing will look more natural this way, because these lines will be somewhat curved: taking the Z for the bump mapped pixels, the unit will take already modified values.

In general, we look forward to NVIDIA and ATI introducing new demos and more correct bump mapping support in games.

As for the performance of the Pixel Shader unit of NVIDIA GeForce4 Ti4600, we discovered no surprises here: the Pixel Shader units of GeForce4 and GeForce3 perform equally fast at the equal working frequencies. And when GeForce4 Ti4600 worked at nominal frequencies, it undoubtedly beat the predecessor:

Accuview and Full-Scene Anti-Aliasing 4xS

Different versions of the full-scene anti-aliasing involving multisampling, 2x, Quincunx, 4x and now one more, 4xS technique, together with textures anisotropic filtering are now called with one single word - Accuview.

However, there is only one thing really new here: 4xS Full-Scene Anti-Aliasing technique.

In fact 4xS technique, which was actually working by NVIDIA GeForce3 as well, is a combination of 2x multisampling and 1x2 supersampling. The final pixel color in case of 4xS is obtained as an average of the two subpixels located one below another, each of which is formed by 2x multisampling. To understand it better, imagine a "2:2" domino piece standing vertically.

4xS technique smoothens the "steps" on the horizontal and slightly inclined lines. We checked it with the help of Serious Sam: The Second Encounter, where we compared the quality provided by 2x, Quincunx, 4x and 4xS techniques from NVIDIA. For a more illustrative example, we also added the results obtained for SMOOTHVISION 6x Quality from ATI:

No Anti-Aliasing	Accuview 2x	Accuview Quincunx
Accuview 4x	Accuview 4xS	Smoothvision 6x Quality

It's true: 4xS technique performed not bad at all when applied to the slightly inclined lines. The outcome was comparable with the quality provided by SMOOTHVISION 6x Quality from ATI, which we considered the today's best solution of the kind. In case of other angles of slope, 4xS also proved very efficient, though the image quality it provided was not so good, in our opinion, as the quality of SMOOTHVISION 6x Quality from ATI. However, the more people, the more opinions...

Proceeding with the Full-Scene Anti-Aliasing, we would like to remind you of one of the multisampling peculiarities: there is one and the same color used for all subpixels (samples), i.e. no matter how many subpixels (samples) are involved into multisampling, the textures or Pixel Shader are applied once for each pixel in a single pass. Therefore, we can't even think about any improvement in texture filtering quality in case multisampling is enabled.

4xS technique is completely different from all other multisampling methods implemented in NVIDIA GeForce3 and GeForce4. The matter is that 4xS uses both: multisampling and supersampling.

In case of the good old supersampling, each subpixel has its own color that is why the texture filtering quality can actually get somewhat better. However, unfortunately, it will happen at the expense of performance, causing very tangible slowdown. Moreover, the menu elements as well as different status bars, which have become an inalienable part of any modern game may appear awfully blurred.

4xS technique offers us a compromise, right? Well, we'll see what this technique provides us in the same Serious Sam: The Encounter with the top level anisotropic filtering enabled:

No Anti-Aliasing	Accuview 2x	Accuview Quincunx
Accuview 4x	Accuview 4xS	Smoothvision 6x Quality

Well, we do see some improvement in texture filtering quality in case of 4xS, but it is so insignificant that you will have to take a really close look to notice it :) which is not what you would be happy to do, we suppose.

However, it is certainly very interesting to find out what influence the use of Full-Scene Anti-Aliasing has on the performance. Let's see how greatly drops the performance in Serious Sam: The Encounter (we tested in Speed and Quality modes, see their description below):

Nothing unexpected happened, all the results turned out just what we had expected them to be. Only in Quality mode 4xS technique runs slower than 4x.

Quality mode differs from Speed mode by anisotropic filtering support, while 4xS differs from 4x - by 2x1 supersampling.

For this particular reason, namely because of the 2x1 supersampling in 4xS, the doubling of the performance needed for proper color calculation appears very tangible when anisotropic filtering is enabled. Moreover, the higher is the level of anisotropic filtering, the more 4xS falls behind 4x. In order to see what's going one when the performance drops in 4x and 4xS modes, we used Lobby test scene from 3DMark2001 benchmark set:

Well, everything is more than evident: the increase in computational burden required by anisotropic filtering results into the same increase in time needed to draw a pixel for 4x and into twice the increase in time for 4xS.

As for the image quality provided by anisotropic filtering on GeForce4, we will dwell on it a little bit later.

Lightspeed Memory Architecture II

Lightspeed Memory Architecture II (LMA II) is a continuation of Lightspeed Memory Architecture implemented in GeForce3.

LMA II includes a 128bit memory controller divided into 4 32bit ones, like the one used in NVIDIA GeForce3, and some other technologies aimed at increasing the efficiency of the graphics memory. They are:

• Z-Occlusion Culling. This method was first introduced in NVIDIA GeForce3. It allows determining if the pixel is visible by its Z-coordinate, before the texturing is done. And if the pixel is invisible in the scene then it will not be processed thus saving time on addressing the frame buffer, Z-buffer and textures.
• Lossless Z Compression. This technology is also implemented in both: NVIDIA GeForce3 and NVIDIA GeForce4. It allows compressing the Z-buffer as 4:1 this way saving quite a lot on data transfer and storage in the Z-buffer.
• Fast Z Clear. This function allows clearing Z-buffer and filling it with the initial Z value before building every new frame. NVIDIA claims that this function was for the first time implemented in GeForce4.
• Memory Auto Pre-Charge. This technology prepares the graphics memory device proactively to pre-charge areas of the memory that are not in use now but are likely to be used in the very near future. NVIDIA claims that it helps avoid forced idling of the GPU lasting up to 10 clocks, when the GPU is actually waiting for the graphics memory to get ready for further action. The wait state in case of Auto Pre-Charge drops down to 2-3 clocks, as NVIDIA engineers claim.
• Quad Cache. NVIDIA GeForce4 features 4 independent buffers (caches). Individual caches store primitive, vertex, texture, and pixel information. These caches are individually optimized for the specific information they deal with.

Well, the list of technologies intended to increase the efficiency of the new NVIDIA core looks really impressive. Now let's see what its practical value is :) There is a very good test, which was some time ago developed by one very good company to demonstrate all the advantages of one very good chip :) We are speaking about VillageMark:

Enabling Z Occlusion Culling and Lossless Z Compression on NVIDIA GeForce4 Ti4600 didn't lead to any noticeable performance drop, which may mean two things: either RivaTuner doesn't know yet how to work with NVIDIA GeForce4, or enabling these functions really doesn't lead to any performance losses.

Of course, it is very pleasing that GeForce4 Ti4600 working at the same frequencies as GeForce3 Ti500 (i.e. downclocked GeForce4 Ti4600) appears faster than the predecessor in all the benchmarks. It means that new features of LMA II architecture are quite useful and ensure that the graphics memory will be used with greater efficiency than in case of the LMA architecture.

nView

nView, a continuation of TwinView ideas, seems to be the most exciting and long-awaited innovation introduced in NVIDIA GeForce4. The new NVIDIA chip from the top price group has finally acquired fully-fledged support of dual-display configurations. Now we have every right to forget about a very unpleasant feeling we had when it turned out that fast and expensive 3D chips from NVIDIA didn't support anything like that, and their cheaper modifications (GeForce2 MX/MX200/MX400) could work with two monitors but failed to provide fast gaming performance.

Of course, ATI's influence here is undeniable: if ATI RADEON 8500/7500 didn't have dual-display support, we could keep dreaming about an alternative solution from NVIDIA for quite a longer while, we believe. However, now we can say thank you to ATI for the dual-display support finally implemented in NVIDIA GeForce4.

NVIDIA GeForce4 chips feature 2 RAMDAC 350MHz, support TV- and digital monitor Outs.

When there are two monitors connected (to be more exact, two displays, i.e. a monitor, and LCD panel or a TV-set), nView allows working in Clone, Zoom, Horizontal Span and Vertical Span modes. Moreover, the primary and secondary devices can be swapped absolutely painlessly, like by all the latest ATI solutions:

You can also set the display orientation properties:

During video playback there is one more working more available: Video Mirror. However, we will not pay special attention to it, as it works just like a similarly called option in TwinView.

Speaking about nView we can't help mentioning the software support of the NVIDIA GeForce4 features implemented on the hardware level. This is where nView Desktop Manager shows its best proving highly functional and convenient when working with nView dual-display configurations.

• Creating and editing profiles
  
  
• Window management and dialog box control
  
  
• Effects and zoom settings
  
  
• Hot key settings
  
  
• Virtual desktops management
  
  

From the functional point of view, nView doesn't yield to HYDRAVISION from ATI. The only remark we could make after working with NVIDIA's sample for a while is about its TV-Out.

According to NVIDIA's statements, there is an enhanced TV-Out unit with HDTV support built into GeForce4 chips. However, the NVIDIA GeForce4 Ti4600 sample we had at our disposal was equipped with an external CX 25871-13 chip from Conexant. Besides, the image quality on the TV screen left much to be desired: ATI based graphics cards can boast a much better TV-Out implementation.

All in all, we look forward to the mass graphics cards on NVIDIA GeForce4 chips to come out, hoping to see a much better made TV-Out on them.

Video Processing Engine, MPEG2 Hardware Decoding

Video Processing Engine is a name for the set of functions dealing with overlays and video processing.

NVIDIA GeForce4 supports all standard functions of the kind, such as motion compensation, color space conversion, etc… Besides, NVIDIA GeForce4 supports 5 horizontal x 3 vertical taps filtering, advanced adaptive de-interlacing during the playback of interlaced video on a non-interlaced monitor, hardware MPEG2 decompression (IDCT: inverse discrete cosine transform). For overlays you can adjust the color and you can also adjust the color settings for each display device separately (Digital Vibrance Color). It also supports HDTV output format.

All in all, VPE offers everything possible for convenient work with video on your PC and convenient video watching on the monitor or TV screen. Unfortunately, the image quality provided by the TV-Out on the NVIDIA GeForce4 Ti4600 sample we had at our disposal left much to be desired, as we have already told you.

So, now that we have got acquainted with all the major and minor innovations introduced in NVIDIA GeForce4 chip, let's have a look at the reference card built on it: NVIDIA GeForce4 Ti4600.

Closer Look: NVIDIA GeForce4 Ti4600 Reference Card

The sample is of pretty big size, is equipped with unusual graphics memory chips, a solid looking cooler, powerful capacitors - in other words, it makes a really great impression:

  

The graphics chip is equipped with a relatively large cooler with NVIDIA logo. NV25 chip, like NV20, is made with 0.15micron manufacturing technology, but features more transistors and higher working frequency that is why it badly needs more efficient cooling:

   

As for the NV25 chip itself, it doesn't have anything special in its appearance. It is only the marking that gives away the today's fastest and most dangerous graphics monster:

The chip works at 300MHz clock frequency, which is the today's highest clock frequency among all 3D gaming chips.

The card is equipped with 128MB DDR SDRAM graphics memory from Samsung with 2.8ns access time. The memory chips feature new BGA package:

This package is probably required by extremely high memory working frequencies and thus the necessity to reduce the EMI and other negative influences. The memory of NVIDIA GeForce4 Ti4600 works at 650Mhz (325MHz DDR) as default. Before today we could achieve these high memory frequencies only by overclocking ATI RADEON 8500 to its utmost. Frankly speaking, we can't wait to try overclocking this way our new NVIDIA GeForce4 Ti4600 :)

The card is equipped with a VGA-, DVI-I and S-Video Outs. The signal for digital monitors is formed by Sil164ct64 chip from Silicon Image, which is very often used on NVIDIA based graphics cards:

The TV-signal is formed by Conexant CX25871-13 chip:

Testbed and Methods

All test were run in the following testbed:

• AMD Athlon XP 2000+ CPU;
• MSI K7T266 Pro2 v2.0 (VIA KT266A) mainboard;
• 2 x 128MB PC2100 CL2 DDR SDRAM by Nanya;
• Fujitsu MPF3153AH HDD.

We used the following software and applications:

• ATI RADEON 8500: driver ver. 6.13.10.6011 for Windows XP;
• VisionTek Xtasy 6964 (NVIDIA GeForce3 Ti500): Detonator 23.11 for Windows XP;
• NVIDIA GeForce4 Ti4600: Detonator 27.20 for Windows XP;
• Windows XP;
• VillageMark;
• Max Payne;
• Serious Sam: The Second Encounter;
• 3DMark 2001;
• Quake3 Arena v1.27.

We tested with the following settings:

• VillageMark:
  texture color depth and frame buffer were set to 32bit. All other settings were set by default.
• Max Payne:
  Quality mode. Image quality set to the maximum. FSAA disabled, anisotropic filtering disabled in order to make the testing conditions equal for the graphics cards based on chips from ATI and NVIDIA. 32bit texture color depth and frame buffer.
  Speed mode. Image quality set to the minimum. 16bit texture color depth and frame buffer.
• Serious Sam: The Second Encounter:
  Speed mode. Image quality set to "Speed", 32 bit graphics modes.
  Quality mode. Image quality set to "Quality", 32 bit graphics modes.
  Extreme Quality mode. Image quality set to "Quality", 32 bit graphics modes. Besides, we also ran the standard add-on "Graphics: Extreme Quality", anisotropic filtering set to the maximum.
  The tests were run on a standard "The Grand Cathedral" demo.

  When we tested NVIDIA GeForce4 Ti4600 reference card we had to add a special line to the graphics auto-initialization file of Serious Sam: The Encounter, so that the gaming engine could recognize NVIDIA GeForce4 Ti4600 graphics chip. We also used the NVIDIA GeForce3 settings for it. We believe, this was the best configuration we could think of here, especially taking into account that the 3D architecture of GeForce4 and GeForce3 chips is almost fully identical.

  If we hadn't done this, the gaming engine wouldn't have recognized NVIDIA GeForce4 Ti4600 based graphics card and would have worked with it as with a "standard one" not involving any of the GeForce3/GeForce4 peculiar features, such as laying 4 textures per clock, etc.

  The TRUFORM technology enabled by this gaming engine as a default setting for ATI RADEON 8500 wasn't remained enabled.

• 3DMark2001:
  All gaming tests were carried out with 32bit frame buffer, 32bit textures and 32bit (24bit) Z-buffer in Pure hardware T&L mode. If the testing conditions changed in synthetic benchmarks, you can read the remarks on the diagrams directly.
• Quake3 Arena v.1.27:
  All test were run with the maximum quality settings, tri-liner filtering and texture compression were enabled. The color depth and frame buffer were set to 32bit.
  We tested in a standard Quake3 Arena 1.27 Point Release Patch demo.

Performance

Synthetic Benchmarks

We have already shown you some of our 3DMark 2001 results above that is why here we will provide only those tests, which we haven't yet paid any attention to:

Well, the situation is very interesting: in 16bit mode the results shown by NVIDIA GeForce4 Ti4600 differ from the possible maximum for this chip only by the measuring error! In 32bit mode (32bit color depth and frame buffer) LMA II provides minor performance losses, so that even at 240Mhz core and 500MHz memory frequency the fillrate by GeForce4 Ti4600 is higher than the fillrate by GeForce3 Ti500.

In the benchmarks testing the bump mapping speed the laurels belonged to NVIDIA GeForce4 Ti4600 due to higher fillrate and Lightspeed Memory Architecture II.

Gaming Benchmarks

The advantage of NVIDIA GeForce4 Ti4600 is simply overwhelming! Dragothic test uses vertex shaders to implement the animation for all the characters, and the number of polygons used in models is very high. No wonder that GeForce4 Ti4600 left all the rivals so far away.

It's a great pity that ATI RADEON 8500 failed to win the lead here. Last time is showed higher results, having left NVIDIA GeForce3 Ti500 behind. However, though not only NVIDIA but also ATI have released the new drivers since then, the situation changed in favour of NVIDIA's offspring.

Here the geometry is not so complex any more. Everything is determined by the fillrate. NVIDIA GeForce4 Ti4600 sample boasting the highest working frequencies and efficient memory subsystem turned out an indisputable leader.

Again we have to admit that the situation for ATI RADEON 8500 changed for the worse. As the practice (and this test as well) shows, high fillrate is not enough for success. The drivers optimization is also of great importance. That is why although ATI RADEON 8500 features the highest fillrate of all, its drivers optimization leaves much to be desires that is why the victory went over to NVIDIA here.

In this benchmark with Pixel and vertex Shaders, NVIDIA GeForce4 Ti4600 again leads the race, which is not at all surprising, we should say: the Pixel and Vertex Shaders speed directly depends on the core working speed, especially since NVIDIA GeForce4 Ti4600 features doubled T&L and Vertex Shaders units. However, the results show that the Vertex Shaders speed is no limiting factor here. The results depend mostly on the fillrate and Pixel Shaders speed.

At least here we can be happy with the performance of ATI RADEON 8500: we see that the results more than doubled compared with what we saw in the previous testing session. Moreover, this performance increase is not caused by the use of faster Athlon XP 2000+ CPU (last time we used Athlon XP 1600+).

The performance in Quake3 Arena is determined by the fillrate and HSR efficiency, that is why NVIDIA GeForce4 Ti4600's leadership here is quite natural :)

In both modes of MaxPayne NVIDIA GeForce4 Ti4600 proves the fastest.

Despite the more advanced T&L unit than that of NVIDIA GeForce3 Ti500 and higher fillrate, ATI RADEON 8500 appeared an outsider here.

Serious Sam: The Second Encounter is a game, which can load to the full extent even a GeForce4 in extreme quality mode :)

However, NVIDIA GeForce4 outperforms all its competitors in all modes, and greater graphics complexity leads only to greater gap between GeForce4 Ti4600 and GeForce3 Ti500. In extreme quality mode, it is not only textures and other special effects that eat up the speed. The objects models are very complex and detailed even if they are located at a long distance, which allows the doubled T&L unit of GeForce4 show its best.

Due to fast running anisotropic filtering, ATI RADEON 8500 proved not bad at all even with more complex graphics involved. It managed to nearly catch up with NVIDIA GeForce4 Ti4600 in the "heaviest" mode.

By the way, the next part of our review is devoted exactly to the fastness of anisotropic filtering and the image quality.

3D Image Quality

Accuview, a new word from NVIDIA, stands for the combination of Full-Scene Anti-Aliasing and anisotropic filtering, the major methods of improving the image quality in the contemporary 3D games. We have already discussed Full-Scene Anti-Aliasing above. Now it's high time we said something about anisotropic filtering.

NVIDIA GeForce4, just like GeForce3 supports 8-, 16- and 32-sample anisotropic filtering. Note that the texture filtering quality will be the same for both chips. To illustrate the statement, let us show you some screenshots from Serious Sam: The Encounter in extreme quality mode:

NVIDIA GeForce4 Ti4600	NVIDIA GeForce3 Ti500	ATI RADEON 8500

As you can see, the quality of anisotropic filtering by NVIDIA GeForce4 Ti4600 and NVIDIA GeForce3 Ti 500 is absolutely the same.

Now let's see how greatly the performance drops once anisotropic filtering is enabled on NVIDIA GeForce4 Ti4600 compared with NVIDIA GeForce3 Ti500 and ATI RADEON 8500:

With anisotropic filtering enabled the performance of NVIDIA GeForce4 Ti4600 dropped down much more than that of NVIDIA GeForce3 Ti500. This is certainly very upsetting, especially bearing in mind that GeForce3 Ti500 already loses almost 50% of its performance when anisotropic filtering is enabled. Only the fact that NVIDIA GeForce4 Ti4600 works are higher frequencies saved the chip from being defeated by GeForce3 Ti500 predecessor.

Overclocking

We would like to remind you that the default working frequencies of NVIDIA GeForce4 Ti4600 are 300MHz for the core and 650MHz (325Mhz DDR) for the graphics memory.

During overclocking we managed to make the card work stably at 330MHz core 740MHz (370MHz DDR) memory. The use of 2.8ns memory chips in BGA packages gives us some hope that the mass graphics cards will boast the same or even higher overclocking potential.

As for the core overclocking on mass pieces, it's very hard to forecast anything now. On the one hand, 300MHz is already quite a lot for a graphics chip made with 0.15micron manufacturing technology, but on the other hand, it is still overclockable, a we managed to increase its working frequency up to 330MHz. So, we think that the mass NV25 based graphics cards will be subject to overclocking just like all the previous generation pieces. :) The only matter to betaken into consideration is cooling.

The cooling of the graphics chip is highly efficient: we wouldn't say that the Nv25 chip on our sample card heated much greater than NVIDIA GeForce3/GeForce3 Ti500. However, no one knows what cooling solutions will be used on the mass graphics cards.

Conclusion

We got the impression that when working on its GeForce4 solution, NVIDIA managed to single out all the weak points of the GeForce3 Ti500 generation and improved all the currently important issues.

The situation, which formed in the market after the launching of ATI RADEON 8500, when NVIDIA GeForce3 Ti500 appeared defeated by the newest graphics cards from ATI, couldn't make NVIDIA happy, of course. That is probably one of the reasons why we see NV25 not as a sign of the coming revolution in 3D graphics but just as a hard-edged response to the competitor.

Look here:

ATI RADEON 8500 unlike NVIDIA GeForce3 supports dual-monitor configurations. Now this support is implemented in GeForce4 as well.

ATI RADEON 8500 can perform hardware MPEG2 decompression. Now GeForce4 also acquires hardware MPEG2 decompression.

ATI RADEON 8500 has a more efficient Vertex Shaders unit than NVIDIA GeForce3. GeForce4 got 2 units of the kind.

ATI RADEON 8500 supports more "advanced" Pixel Shaders than NVIDIA GeForce3: version 1.4 against version 1.1. By GeForce4 NVIDIA squeezed all juices out of the old architecture to announce Pixel Shaders version 1.3 support.

ATI RADEON 8500 boasts HyperZ II, which is more efficient than Lightspeed Memory Architecture of GeForce3. GeForce4 now boasts Lightspeed Memory Architecture II.

At last, ATI RADEON 8500 works at higher core and memory frequencies than GeForce3 Ti500. GeForce4 Ti4600 is now working at even higher frequencies.

All in all, we have every right to state that NVIDIA GeForce4 has almost no bottlenecks, and its performance is so high that the other today's graphics chips will not be able to get even close to it.

Well, some of you may be upset that no 3D revolution burst out, however, revolutions are not always a good thing.

Firstly, fully-fledged support of DirectX8 in games hasn't become mass and a new revolution could possibly harm the coming of DirectX8 and shaders into "public".

Secondly, the launching of new graphics cards will cause the prices of the already available solutions to drop, which will make the graphics cards with hardware support of DirectX8 more affordable for the users. That is why those people who decide not to spend too much money on NVIDIA GeForce4 based graphics cards will be pretty happy with a GeForce3 Ti500/Ti200 or ATI RADEON 8500, as they will lose only a bit of performance and nothing else. And those who buy an NVIDIA GeForce4 can be absolutely sure that their graphics card is beyond any competition.

The launching of the chips supporting DirectX9 is now moved to a later day, and then the new wars will break out for sure: between NV30 and R300. Maybe by that time a wonder happens and new PowerVR/STM and S3 Graphics chips crop up. Then the battle will turn out even more exciting.

And in the meanwhile we can only look forward to the first mass graphics cards on NVIDIA GeForce4 to come, and to a new investigation of the mysterious GeForce4 MX, a low-cost solution for those who want to keep pace with the time.