Life is going its natural course and the new graphics cards keep on coming into the world. We want to remind you that an incredibly great number of chipsets and graphics cards are planned for this spring. Some of them have already turned up in the market, and some of them were exhibited at Comtek'99. Among those, which attracted our attention at the computer show were S3 Savage4 (Pro) chipset and Diamond Stealth III S540 graphics card based on it.

Here are the main specifications of Savage4 Pro:

	S3 Savage4 Pro	ATI Rage128	Nvidia Riva TNT2	3dfx Voodoo3
API support	Direct3D, OpenGL	Direct3D, OpenGL	Direct3D, OpenGL	Direct3D, Glide/OpenGL
Core frequency, MHz	143	100	125-183	143-183
Memory frequency, MHz	143	110	125-183	143-183
RAMDAC, MHz	300	250	300	300
Max supported memory, MB	32	32	32	16
3D Truecolor (32bit) rendering	yes	yes	yes	no
3D maximal resolution
- Highcolor (16bit)	1920x1440	1600x1200	2048x1536	1600x1200
- Truecolor (32bit)	1920x1440	1600x1200	2048x1536	no
The number of rendering pipelines	1	2	2	1
Fillrate, mln pixels per second	143	200	300	143-183
Processing power, mln polygones per second	8	4	8	6
Larger (1024x1024) texture support	yes	yes	yes	no
AGP support
- DiME	yes	yes	yes	no
- AGP 2x	yes	yes	yes	no
- AGP 4x	yes	no	yes	no
Z buffer accuracy	24	32	24	16
Per pixel MIP-mapping	yes	yes	yes	yes
Auto MIP-mapping	yes	yes	yes	yes
Tri-linear filtering
- Single pass	yes	yes	yes	yes
- Approximation	no	no	yes	no
Anisotropic filtering	yes	no	yes	yes
Multitexturing	yes	yes	yes	yes
Anti-aliasing
- edge	no	no	no	yes
- full-scene	yes	yes	yes	yes
Fog	yes	yes	yes	yes
OpenGL support	ICD	ICD	ICD	Miniport (wrapper)

Below you can see the pictures of the graphics card and memory chip Diamond Stealth III S540:

As we can see this card is based on S3 Savage4 Pro chipset, which differs from Savage4 only by higher working frequencies (143MHz compared to 125MHz Savage4). The card has 16MB SDRAM 7 ns of memory placed on both sides of the card (4 modules on each side). Unlike plenty of other graphics cards, Diamond Stealth III S450 has a special plug, which allows to display information on LCD-monitors.

Potentially, Diamond Stealth III S450 may have a TV-out because there are some specially designed locations for a chip and TV-out connector on the card. There is also a black radiator stuck to the chipset.

Since we've managed to get only an experimental sample for benchmarks we can hardly say anything concerning its package set. The graphics card has fallen into our hands absolutely "naked", even without the drivers. (However, thank goodness, we could download them from the Internet).

Frankly speaking, we were expecting the graphics cards on Savage4 to appear pretty soon since their terrific image quality had been the talk of the town. Besides, not so long ago S3 made several announcements about extremely high performance of this chipset. As is known, Savage3D and Savage4 can operate with compressed textures, in other words they are compressing and decompressing them "on fly". This allows to get a very high quality image - even a photographic one, containing larger textures - without any tangible speed reduction. As a result the 3D-scene quality increases greatly from the point of its visual perception.

And now let's pass over to the benchmarks themselves. The testing system was configured as follows:

• ASUS P2B-B (440BX) mainboard;
• Intel Pentium II 450 CPU;
• System memory 128 MB PC-100;
• Nokia 447Xav (17") monitor;
• Windows 98.

A few words about the software. The drivers seem to be a bit raw yet (even S3 haven't yet made the reference drivers) because they wouldn't get installed for quite a while (we actually failed to install them through install.exe). And finally after having resorted to manual installation through the INF-file our sufferings were paid and we got an additional page in the display properties list.

The drivers have a profile structure, which means that they form a separate profile for each running application. So, no matter haw many programs you run (besides the games): each of them will appear as a profile with its own properties set. It struck us as really funny when the system suggested enabling fog support for Photoshop, for example. The properties set can't boast much and is relevant only in terms of Direct3D for some reasons. (As you can see from the screenshot OpenGL and Advanced options are inaccessible). This is also very likely to be caused by the unfinished drivers. Nevertheless, it is very convenient that you are allowed to define independent set of parameters for each game, for example, you can enable multitexturing only where supported.

All the profiles can be deleted and saved as a ini-file. Besides, the already composed profiles can be imported from the existing ini-files.

Well, that seems to be about all concerning advanced display properties with the only exception of color gamma correction.

The next question, which is worth dwelling on, is the ability of the card to work at frequencies exceeding those stated by its specs list. We would like to remind you that the nominal frequency for S3 Savage4 Pro makes 143MHz for both: core and memory. A well-known utility S3 Tweak used by numerous Savage3D owners for changing the settings of their card and for overclocking can be also applied to Savage4.

Its newer version 2.0 supports this chipset as well. However, the restrictions to 150MHz for core frequency and to 143MHz for memory frequency turned out an insurmountable obstacle. (Though we laid ourselves out trying to find something in Registry and S3 Tweak properties set, which could be of any use.)

Our fruitless efforts made us think that the drivers contain a certain delimiter, which we unfortunately failed to detect. However, we do not shrug off some other possible reasons restricting overclocking. That is why all the benchmarks results given further on were obtained only at nominal frequencies equal to 143MHz for core and memory.By the way, the newest version of PowerStrip (2.41.07) doesn't work properly with Savage4 though it is supposed to.Moreover, it not only eliminates the possibility to change the frequencies - it even displays patently false values.

S3 Tweak utility allows to operate with main 3D functions but unfortunately a number of options is not carried out in the existing drivers that is why your enabling or disabling them won't lead to any changes.

Now let's consider the performance of this card in 2D-graphics. As far as the image quality is concerned, Diamond Stealth III S450 is lagging far behind compared to all the graphics cards we have tested recently. We would position it a bit behind Riva TNT because at 1280x1024 the picture appears oppressively washed out though RAMDAC frequency is 300MHz. However, it is quite possible that those of you who have a 15" or 17" monitor won't regard it as a serious drawback because the resolution equal to 1024x768 and up is simply not needed in this case.

So, and what do we have here? Aha! It's the speed achieved in 2D. The corresponding diagrams posted below represent the results of WinBench99 at 1024x768 in 16-bit color:

Well, there is nothing to be happy about. The speed of S3 Savage4 in 2D is lower than that of the two new chipsets.But still Diamond Stealth III S540 is a purely gaming card that's why some of you may argue about the importance of high 2D speed requirements.

Then we will pay attention to the performance of the new graphics card in 3D-graphics, and first of all in terms of its speed. The performance of Diamond Stealth III S540 will be compared with that of the cards based on the following chipsets:

• nVidia Riva TNT2 (ASUS AGP-V3800);
• 3dfx Voodoo3 2000;
• 3dfx Voodoo3 3000;
• ATI Rage 128 (ATI Rage Fury).

We will offer you the results of two types of testing, as we usually do it. Firstly, while the CPU frequency (Intel Pentium II) is fixed as 450MHz, we measured the speed achieved at the resolutions from 640x480 to 1280x1024. To do this we applied three different tools:

• 3D Mark MAX (DirectX);
• Quake2 (massive1.dm2) (OpenGL);
• Incoming (DirectX).

Secondly, in 3D Mark MAX we got the speed values obtained at three different frequencies of the Intel Pentium II CPU (450, 300 and 233MHz correspondingly). This let us determine the dependence of the graphics card performance on the CPU frequency.

So, how can we evaluate the performance of Diamond Stealth III S450? Let's start with 16-bit color regime. Judging by the results of the most benchmarks in Direct3D the Savage4 card proved to be among the coolest gaming accelerators (such as Voodoo3 and Riva TNT2). However, this cannot refer to OpenGL. A relatively incomplete ICD OpenGL by Savage4 makes itself felt. Besides, there is one more important point that should be mentioned: presently OpenGL doesn't support 32-bit color.

If we have a look at the diagram showing the performance drop occurring when we shift from 16- to 32-bit color,

We will definitely notice the dual character of the results obtained. The one hand Savage4 performed beautifully at lower resolutions (below 1024x768) and gained the lead. But on the other hand its speed has significantly dropped at 1024x768 (though it is not so great as by Riva TNT). It can be caused either by the insufficient CPU capacity or again by raw drivers.

However, the speed in 32-bit color is quite high. Let's discuss the performance of Diamond Stealth III S540 here.And what have we got? Brilliant results in 3DMark 99 MAX and rather mediocre achievements in Incoming. Actually, this difference seems to be justified. To make it clear we suggest telling you a bit about multitexturing.

Most of you are definitely aware that modern chipsets have two TMU processors working with textures and one processor for the buffer receiving the already textured pixels. Thus multitexturing allows to superpose two textures over one pixel per time step, which is supposed to increase the texturing speed and hence the general performance.

And right here the processors are distinguished relative to the ways they treat the textures. Such graphics chipsets as Voodoo2 and Voodoo3 have only one rendering pipeline but two texture processors (TMU). In other words, in normal conditions one TMU is standing idle, while during multitexturing both of them are busy and the superposing of two textures over one pixel occurs simultaneously. Voodoo2 and Voodoo3 demonstrate the same fill rate in both cases with multitexturing and without it. And as a result the final performance of Voodoo2 and Voodoo3 almost doubles when multitexturing is enabled.

Riva TNT, Rage 128, Riva TNT2 graphics chips are also equipped with two TMUs. But these cards utilize them in an absolutely different way. Without multitexturing rendering is carried out through two pipelines simultaneously, which means that one texture can be superposed over two pixels per time step. Enabling multitexturing changes the whole picture. Now two textures can be superposed over one single pixel per time step. So, multitexturing allows to treat only one pixel per time step that is why the fill rate reduces twice. And this practically means that both texturing regimes by Riva TNT (2) and Rage 128 should provide the same performance level.

In this respect a very natural question comes to mind: which regime is better if the game supports multitexturing? What is more important: to faster operate with pixels (two per time step) and then to return to next texture superposing (usual regime) or to sacrifice high speed (one pixel per time step) for the sake of two textures superposition (multitexturing)? The answer can be found only by experiment. For instance, Riva TNT showed a 15-20% performance increase with multitexturing.

But back to Savage4. In terms of multitexturing it is very much like Voodoo3, i.e. there are two TMUs but only one is active in usual regime.

There are two gaming episodes in 3DMark99 MAX benchmarks set: Game1 and Game2, which contribute most of all the rest to the final result. But Game1 (as well as Incoming) doesn't use multitexturing, and Game2 does. The performance of Savage4 in Incoming turns out lower than that of Riva TNT2. But as soon as we pass over to 3DMark99 MAX Game2 where multitexturing takes its turn, Savage4 proves to be much cooler than its rival. So, here we can draw a preliminary conclusion, that this algorithm is very effective with Savage4 especially in 32-bit color. And the conclusion is preliminary because we failed to see the same effect in Quake2, which also supports multitexturing, because of the drivers incompleteness.

A small bit about filtering. To eliminate the pixelization of nearby objects there is a special means called bilinear filtering, which evenly blurs the image. However, it appears insufficient for a good quality picture. To make the objects moving away look more true to life the so-called MIP-levels were introduced. They allow to display the same texture with various levels of detailization. The farther is the object from our eyes, the less clear are its contours. But unfortunately the boundaries between the MIP-levels are very coarse and thick and can be easily noticed and hence spoil the whole picture. Bilinear filtering is unable to combat this effect that's why tri-linear filtering is activated. However, the new chipsets still have some problems supporting it. Almost all manufacturers announce tri-linear filtering absolute support of their products. But in fact only few chipsets provide the required built-in support for this technology (Matrox G200, 3dfx Voodoo2 and Voodoo3). All the rest including all nVidia chipsets content with approximation, a sort of emulation. Though the difference is noticeable only in the image quality of the real games.

What did we see in case of Savage4? Tri-linear filtering proved to be on a very high level, and the speed stayed almost the same all the time: no drops! Comparing the results to Riva TNT2 we can state that Riva suffered a 16-18% speed decrease as soon as the tri-linear filtering had been enabled. Judging by all these facts we may declare a very good filtering optimization of Savage4.

So, a very high speed in 32-bit color, as well as brilliant performance in 16-bit color, which positions Savage4 among the coolest new 3D-accelerators, contribute to a perfect reputation of this graphics card. The only drawback - low performance in OpenGL - may be caused by imperfect drivers or by a temporary lack of some important functions. That's why we have every reason to hope that the situation will change for the better in the near future.

Since the graphics cards based on S3 Savage4 cost less than $150 for a 32MB-version this product may cause a very cut-throat competition to the already available Voodoo3 and Riva TNT2.

And now let's talk a bit about the 3D-graphics quality and larger textures support of Savage4.

First of all, we would like to briefly present the results of 3DMark 99 MAX. There are two points worth mentioning here.

1. Perfect image quality in Game2:

Reference Image	S3 Savage4 Pro
nVidia Riva TNT2	3dfx Voodoo3

You can see that the washed out effect typical of Voodoo3 and occurring by Riva TNT2 is simply absent here! Well, the quality even seems to have surpassed the reference image! :)

2. Excellent support of subpixel correction.

Reference Image	S3 Savage4 Pro
nVidia Riva TNT2

Here too Savage4 left all the others far behind.

As for Direct3D games - no remarks here. The image is of really high quality. Unfortunately, we have to admit that the drivers lack anti-aliasing support but we hope it will be introduced later.

It is also interesting that the support of so greatly praised S3TC (texture compression) technology is included only perforce. What we mean here is that it turns out absolutely impossible to check the accelerator performance with the enabled and disabled S3TC in the same applications. And now S3TC is enabled only in case some large compressed textures are used (in a couple of Unreal levels, for instance). In all other cases the compression can't be enabled.

OpenGL in games also showed a rather high image quality. We have on purpose installed a demo-level from Quake2 with larger textures:

But the most fantastic image quality appears in API S3Metal developed directly for chipsets supporting S3TC. Though S3Metal is now actualized only in a special version of Unreal, the image quality surpasses all the expectations.

The screenshot below is made from S3Movie level where most textures are none other than a live video. Besides, there is also something like holograms. Here is what one of these pictures looks like from a distance:

And here is a picture with a hologram. Pay attention to the legibility of the text:

We have also checked a famous EgyptF for Unreal: a room with numerous wall pictures. Below, you can see one of the walls:

Note, that the texture clearness doesn't get worse while approaching the surface:

In case of any other card the utilization of such small textures as in this case should undoubtedly result into an awfully washed out image because of the bilinear filtering.

And to conclude with - another beautiful screenshot from the same level:

Well, we have to admit that no other accelerator has ever demonstrated such impressive results! But, we have to put a fly into the ointment: a huge lag (the image changing can't catch up with the player's actions) makes these Unreal levels no fun for playing. We hope that this lag will disappear with the time when the drivers are improved.

The conclusion on 3D will be very short: it's great!

Summing up we would like to say that Diamond Stealth III S540 graphics card performed beautifully and is comparable to all the new graphics cards. 3D quality is perfect. And as for a comparatively low speed in 2D, it is more than compensated by the purely gaming aims of this product together with its relatively low price, which makes it available for most 3D admirers.

Besides, we feel like mentioning once again that all we had was just a sample without proper drivers. That's why we shouldn't make any final decisions unless mass production of Savage4-based graphics cards makes a start.