Today, on October 1, NVIDIA officially announced a new graphics chip family aka Titanium. The rumors about NVIDIA preparing something new have been circulating around the web for a considerable while. Besides, the season, the fall, is really suitable for new product announcements… The most optimistic forecasts predicted the launching of NV25, a new progressive graphics core, which was destined to make another revolution in the graphics market, and NV17, also known as GeForce3 MX, which could bring some GeForce3 features to those guys who don't feel like spending their money on a fully-fledged GeForce3. However, none of these came true. Instead there appeared a new Titanium family. So, what did it bring us?

The fastest and the most expensive chip in the Titanium family is GeForce3 Ti 500. Due to the working frequencies increased up to 240MHz/500MHz it can be regarded as none other but a faster GeForce3 modification and is supposed to inherit the leading position from the current GeForce3. Of course, the developers couldn't help taking advantage of the GeForce3 chip huge overclocking potential and this is exactly how GeForce3 Ti 500 came into being, having postponed the introduction of the revolutionary NV25.

GeForce3 Ti 200, which has stepped in for the so long awaited NV17 or GeForce3 MX, luckily turned out to feature only reduced frequencies, having retained the entire GeForce3 architecture as it is without any changes. The only difference from the "regular GeForce3" is the working core and memory frequencies equal to 175MHz and 400MHz correspondingly.

GeForce2 Ti represents a faster modification of the "regular" GeForce2 Pro. Maybe it saw the light because NVIDIA was not very much inspired with the poor sales of the GeForce2 Ultra equipped with expensive 4ns memory chips and decided to launch a compromising solution with the 250MHz core and 400MHz memory frequency. This way, GeForce2 Ti based graphics cards appeared equipped with the same memory as GeForce2 Pro but will be faster than the latter due to higher core clock frequency.

Here is a chart showing how NVIDIA itself positions the new Titanium solutions:

Well, if GeForce3 Titanium 200 really costs twice as little as GeForce3 Ti 500, it will appear much closer to the Mainstream market segment rather than to the Performance segment. Especially when the solution passes through the "novelty stage" and the card prices get significantly lower than on the launching day.

To tell the truth, it seems quite doubtful that GeForce3 Ti 500 will cost the same sum of money as GeForce3, and GeForce2 Ti - the same money as GeForce2 Pro. We don't think it will make any sense to undermine the sales of the already existing products, so we expect the prices of the new solutions to be somewhat higher.

Actually, when we think how easily NVIDIA gave birth to two new products based on one and the same GeForce3 architecture, it reminds us of the GeForce2 MX story, when we got GeForce2 MX 200 and GeForce2 MX 400 chips. However, if in case of GeForce2 MX family the performance difference between the MX 200 and MX 400 was quite tangible, now Ti 200 shouldn't be too far behind Ti 500. In other words, very soon you might be able to get a graphics card yielding a real bit in performance to the today's leader for a considerably lower price.

Anyway, this review should help us to find out if all this discussion makes any real sense or not. And before we start talking about the newcomers, actually, we would like to mention one more innovation in the graphics world: the new Detonator XP driver. According to NVIDIA press releases, this driver is intended specifically for the Titanium family. Of course, it also works with the existing cards and provides a noticeable performance increase due to better optimization. But the most important thing about it is the fact that NVIDIA has finally implemented in this driver the support for two hardware features of the GeForce3 core, which we haven't yet seen. Since these features are of really great interest to all of us, we suggest taking a closer look at them before passing over to the actual graphics cards.

Shadow Buffer, 3D Textures

Due to Shadow Buffer support NVIDIA GeForce3 core can create real-time shadows.

At first glance, it seems that the game is not worth the candle: what's the use of inventing new technologies and algorithms if the realistic lighting could be added with the use of traditional lightmaps? As you remember from our previous articles, we showed very clearly that the application of certain lightmaps allows introducing significant changes to the 3D scene by loading the graphics card just a little bit more than usual (it needs to lay one more texture, i.e. the lightmap). We are simply used to the way the scenes with applied lightmaps look that is why we do not notice the changes. Just as a quick reminder, take a look at a couple of screenshots from Quake3 Arena. There is no need to point out, which one is taken with the disabled lightmaps, we believe:

Even though lightmaps are pretty simple and easy to apply, they are still not free from some drawbacks. The most unpleasant drawback of this solution is the fact that lightmaps are all static. In other words, it means the following. Take that the game developers created, for instance, a room in such a way that a table, a chair and a bed cast certain shadows from the ceiling lamp onto the floor. So, if you take away the chair, break the table and fire a grenade into the bed, the shadows will still remain on the floor. The matter is that they have been preliminarily calculated by the game developers and included into the lightmap for the floor of this particular room. No doubt, that creating real-time shadows will help avoid confusions like that.

Besides, real-time shadows can be also used to add more realism to the dynamic lighting effects. Imagine a dark damp crypt, where the poor player, locked in the corner, tries to fight several monsters at a time with a rifle, for instance. The flashes from the firing rifle light a small part of the stone floor, the player's face and the monsters' claws. The gamer is scared, the tension is growing… And can you imagine how far more impressive could the whole scene look if in the rifle flashes you could see not only a lonely circle of the stone floor but also some frightening monster shadows on the walls?! We should admit that so far the developers have paid a lot of attention to the shadows cast by the player and some selected surrounding objects. Of course, it adds a lot to the realism of the game. Take, for instance, Blade of Darkness shot:

However, if you take a closer look at the screenshot, you will notice that the torch casts a very clear-cut shadow, as if it were not a torch but a halogen lamp over there. The shadow from a torch like that should be washed out, but it will definitely require much more complicated calculations, which pushed the developers to a compromise: a clear-cut shadow is better than nothing. You will see the same thing in many other games. But we sincerely hope that this problem will be eliminated very soon now. Hardware Shadow Buffer of NVIDIA GeForce3 allows creating much more life-like washed-out shadows.

At last, Shadow Buffer lets the objects cast shadows over themselves, which has always been unacceptable for real-time modeling because it requires great computing capacities if emulated on the software level.

Well, now it's high time we examined the hardware implementation of the GeForce3 Shadow Buffer. For our investigation we will take a scene with an object casting shadow over itself and another object nearby:

So, let's find out what things GeForce3 needs to do in order to generate realistic shadows.

At first, GeForce3 designates an area in the graphics memory as a Shadow Buffer. In fact, Shadow Buffer is a two-dimensional pixel array (as a logical notion) or a memory area (as a physical notion), the same as a frame buffer or Z-buffer.

Then the scene is modeled in the Shadow Buffer, but in a different reference frame: GeForce3 GPU renders the scene from the point of view of the light, as if the viewer were located at the same position and looking in the same direction as the light:

The scene is built in the Shadow Buffer following the same algorithm according to which it is built on the screen, though this algorithm is much simpler: GeForce3 stores only the depth information, or Z-value, for each pixel (i.e. how far the pixel is from the light source) in the Shadow Buffer.

Provided the light hitting the closest non-transparent object cannot go any farther, the smallest Z-values for each pixel are more than enough to ensure that the Shadow Buffer represents only the objects that are closest to the light. If the smaller values stored in the Z-buffer are marked with darker colors and larger values - with lighter colors then we will get the following picture:

Well, the Shadow Buffer is built, everything is ready for further scene creation. It is done in an almost ordinary way, but with an additional calculation carried out for each pixel. The calculation should help find out how far from the light source the part of the object shown by this particular pixel is. If the distance from the considered pixel to the light source (the Z value) turns out greater than the value stored in the Shadow Buffer, then it means that the light coming from this source got stuck somewhere earlier, on the way to the pixel, and this part of the object shouldn't be lit. And if the Z value of the pixel appears equal to (or very close to equal to) the Z value stored in the Shadow Buffer, then the pixel is lit and the light ray doesn't go any farther.

In order to reduce the negative influence of some inevitable errors resulting from inaccurate calculations and to make the shadows softer and smoother there is a proprietary shadow algorithm, according to which the multiple Z-values are compared not only to the single Z-value taken from the Shadow Buffer but to a number of nearby values as well. As a result, the outcome of this comparison appears not just one out of the two possible responses, such as "shadowed" and "not shadowed", but a number of shadow levels equal to the number of nearby values taken from the Shadow Buffer. The maximum number of values that can be taken makes 256, which allows obtaining up to 256 discrete levels of shadow to create soft edges for shadows:

This beautiful solution allows killing two birds with a stone. Firstly, it reduces the risk of errors in shadow calculations, and secondly, it allows creating soft shadows.

Of course, all these extra calculations will tell on the gaming performance, but how great the performance drop will be is a really interesting question. Especially, if we want to have not just a single light source, but 5, 10 or even 50 different light sources in our scene. It means that the Shadow Buffer will need to be recalculated 5, 10 or even 50 times.

On the other hand, every time we create the Shadow Buffer, no texturing is involved, therefore, the graphics memory is not loaded too heavily. Moreover, the Shadow Buffer size can be set depending on the shadows quality you'd like to get in the end. So, sacrificing a little bit of shadow quality may help to reduce the Shadow Buffer size and hence save the performance.

Anyway, we'll see. Currently this technology isn't yet implemented in the existing games, but there is the upcoming Doom III. Judging by the demos we have seen, this game should boast a really great implementation of shadows and lights.

As for the next new feature we mentioned in the very beginning, namely 3D Textures, everything is much easier to explain here. These are regular textures created with one more dimension involved, that's it:

Now you can set the object's texture not only in a two-dimensional reference frame, but also in a three-dimensional one. In this case, the sectional view of the object's model with the given 3D texture will look absolutely real:

3D textures suit excellently for complex environmental effects, such as clouds, uneven fog, volumetric lighting, etc.

Unfortunately, 3D textures are really huge. See yourself: a regular 32bit texture 512x512 requires 1MB of memory, which is quite OK for today. However, a 3D texture 512x512x512 will need 512 times more memory, that is 512MB, which is hardly acceptable for any existing graphics accelerator. Even the 3D texture compression technology of the GeForce3 will allow to reduce the size as 1:8 at the maximum, so that this giant texture will now need 64MB of memory to be stored, which is still too much. Nevertheless, not very large 3D textures can be used to create different atmospheric effects, especially together with bi-linear and tri-linear filtering GeForce3 provides for 3D textures. However, since the textures have become three-dimensional, the filtering modes will now be called tri-linear and quadri-linear (what a word, eh? :)

Well, let's have patience and wait for the new features to find their way in games, and now we are ready to go for the NVIDIA newcomers: Titanium graphics solutions.

Closer Look

GeForce3 Ti 500 reference card at first glance looks absolutely like GeForce3:

However, there are still some differences: the card layout near D-Sub connector has been changed. Take a look:

GeForce3


GeForce3 Ti 500

As we have already said, the card is based on the same NVIDIA GeForce3 rev. A5 chip, despite the new name it has now:

However, due to high working frequency of the graphics memory, the card acquired faster memory chips. As you may see, it is equipped with 64MB graphics memory by ESMT with 3.5ns access time:

All the other peculiarities of the new graphics card are absolutely the same as by the regular NVIDIA GeForce3.

Unfortunately, we didn't have any NVIDIA GeForce3 Ti 200 based cards in our testlab when we were working on the review. That is why in order to get some idea of what GeForce3 Ti 200 actually presents like we downclocked GeForce3 Ti 500, so that its core worked at 175MHz and memory at 400MHz, like by the real Ti 200.

We resorted to the same method in order to get a GeForce2 Ti, namely we tested the GeForce2 Pro with the working frequencies set to 250MHz for the core and 400MHz for the memory. Anyway, we nevertheless managed to get a picture of the real GeForce2 Ti. Here it is:

Testbed

The graphics cards were tested on the following testbed:

• AMD Athlon 1.4GHz CPU (133MHz FSB);
• ASUS A7V133-C (VIA KT133A) mainboard;
• 256MB NCP PC133 SDRAM;
• Fujitsu MPE3084AE 8.4GB HDD.

We used the following software:

• Windows 98 SE build 4.10.2222 A;
• DirectX8.0a;
• Quake3 Arena v1.27g.

We used the new Detonator XP driver for the Titanium graphics solutions.

Performance

Since the performance of the new Titanium graphics cards in different applications is quite predictable and doesn't differ too much from that of the existing solutions, we believe it will be more than enough to focus on Quake3 Arena and 3DMark 2001.

In Quake3 Arena we set the image quality to the maximum, enabled tri-linear filtering and disabled texture compression. The results obtained look as follows:

In 3DMark 2001 for 32bit color modes we selected 32bit textures and 32bit Z-buffer, while for 16bit mode we chose 16bit textures and 16bit Z-buffer. Here come the results for High Detail tests:

Well, as we have told you a few minutes ago, the results appeared pretty predictable. GeForce3 Ti 500, GeForce3 Ti 200 and GeForce3 run neck and neck, and it looks as if GeForce3 Ti 200 were much closer to GeForce3 Ti 500 than GeForce2 MX used to be compared with GeForce2 GTS.

As for GeForce2 Ti, it proved tangibly faster than GeForce2 Pro only in 16bit mode. In 32bit mode the performance difference between these two solutions is nearly negligible, because the memory frequency remains unchanged.

Overclocking

The graphics cards based on the new NVIDIA GeForce3 Ti 500 are equipped not with the 3.8ns graphics memory like the regular GeForce3, but with 3.5ns memory. That is why the maximum working frequencies of the GeForce3 Ti 500 based graphics cards, at least those of the graphics memory, should be not lower than those of GeForce3.

As for the chip frequency, it's quite hard to make any suppositions here. However, anyway, the GeForce3 Ti 500 chips should definitely work fine at 240MHz, unlike the GeForce3 chips working guaranteed at 200MHz. That is why the Ti 500 core should undoubtedly reach higher frequencies when overclocked.

As it came to practice, all our guesses came 100% true. The maximum working frequencies at which the card ran stably made 270MHz for the core and 590MHz for the memory. Of course, it is much higher than 250MHz/560MHz some regular GeForce3 pieces managed to reach, but if we take a look at the increase percentage from the nominal then GeForce3 will still retain the leading position here.

Nevertheless, we haven't yet seen any other NVIDIA based graphics solutions, which could reach 270MHz/590MHz without even minor hardware modifications. And it means that if you consider yourself an enthusiastic user, then it's high time you started thinking about getting a new graphics card. And if you are fond of extreme overclocking, then we would advise you to wait a little bit for our new article on extreme overclocking of NVIDIA's newcomer. :)

Well, let's see what performance growth higher core and memory frequencies can grant the new Ti 500 compared with the GeForce3:

Hm... No doubt, GeForce3, which ahs now turned into GeForce3 Ti 500, still has a huge hidden potential. Therefore, we won't be surprised to see some Ti 600 or Ti 500 Golden Sample one day :) The cool observation we managed to make about these graphics solutions is that at resolution set to 1024x768 and lower Athlon 1400MHz appears the bottleneck slowing down the system. Look: the performance of all the testing participants is almost the same in 1024x768.

Conclusion

NVIDIA did a really smart thing having announced two new products based on the GeForce3 architecture: GeForce3 Ti 500 and GeForce3 Ti 200.

The first one, GeForce3 Ti 500, takes advantage of the GeForce3 performance potential and appears a new leader in the today's graphics market. Surely, it will always find its admirers among the "extreme users". And GeForce3 Ti 200 has every chance to replace the so beloved GeForce2 MX.

GeForce2 Ti doesn't show any significant performance improvement compared to the GeForce2 Pro and is very unlikely to be as attractive for the customers as GeForce2 Ti 200, so we wouldn't dare predict its fortune.

Anyway, we have to wait a little bit before drawing final conclusions, and when the "novelty stage" passes by and the prices get more or less stable, we will make our verdict…