November'2001 3Digest - Dual VSA-100 (3dfx Voodoo5 5500)

By Andrew Worobyew

This tandem is presented by 3dfx Voodoo5 5500 64 MBytes, AGP videocard.

Features:

• Two 3dfx VSA 100 chips, 166 MHz clock rate;
• Two 32 MBytes SDRAM memory banks in four Hyundai 6ns microchips each, 166 MHz clock rate, 128-bit bus;
• Peak speed of scene drawing in multitexturing mode is 333 megapixels/sec and 666 megatexels/sec.

Overclocks to 180/180 MHz (see summary diagrams).

For the 20th of November'2001 the latest version from x3dfx is 1.08.02.

To all appearances there will not be any drivers support from ex-3dfx from now on. This verifies the moral death of the whole Voodoo3/4/5 line. The analysis of tweaked drivers from various enthusiasts has shown the absense of anything new, except for just few. So this section will be archived next month.

Speaking about 1.07.* drivers from x-3dfx, I shall mark that they turned out to be rather good and sometimes even more productive than 1.07.00 from 3dfx, therefore we used them for calculating usability coefficients. The latest 1.08.02 version cause a perfomance falloff. For FSAA perfomance see the summary diagrams.

Below are this card screenshots and the reference ones taken on NVIDIA GeForce3.

Games / Cards	1	2	3	4	5	6
3dfx Voodoo5 5500
NVIDIA GeForce3

Games / Cards	7	8	9	10	11	12
3dfx Voodoo5 5500				-
NVIDIA GeForce3

Games / Cards	13	14	15	16	17	18
3dfx Voodoo5 5500						-
NVIDIA GeForce3

Games / Cards	19	20	21	22	23	24
3dfx Voodoo5 5500	-	-		-
NVIDIA GeForce3

Games / Cards	25	26
3dfx Voodoo5 5500
NVIDIA GeForce3

Game list:

1. Return to Castle Wolfenstein
2. Unreal2 - for GeForce cards, Unreal - for all other
3. Comanche4
4. No One Lives Forever
5. Serious Sam
6. Colin McRAE Rally2
7. Red Faction
8. Rally Trophy
9. Sacrifice
10. Venom
11. Need For Speed V (Porshe 2000)
12. Real MYST
13. Blade of Darkness
14. Giants
15. 3DMark2001: Game1
16. 3DMark2001: Game2
17. 3DMark2001: Game3
18. 3DMark2001: Game4
19. 3DMark2001: Dot3
20. 3DMark2001: EMBM
21. 3DMark2001: Vertex Shaders
22. 3DMark2001: Pixel Shaders
23. Black & White
24. Undying
25. Anachronox
26. Max Payne

There's some criticism about 3D quality. I must note the fact of absence of true trilinear filtering, that was officially confirmed by 3dfx. In the articles about 3dfx Voodoo3 2000/3000, I have already noted that trilinear filtering has disappeared in new drivers.

And there's a confirmation of the above in Quake3 with r_colorMIPLevels "1" option:

NVIDIA GeForce2 GTS shot is to the left, 3dfx Voodoo5 5500 - to the right, Voodoo4 4500 produces the same.

So we must state that 3dfx Voodoo4-5 cards don't support trilinear and anisotropic filtering. But there are some positive moments like the appearance of correct approximation in 1.07.02 version from Iceman. It doesn't differ much from real trilinear filtering, and doesn't cause perfomance falloffs.

The latest Comanche4 game looks like horror:

Let's take a look at LOD Bias settings. It is set to 0 by default. Let's observe what happens after setting it to boundary (-2,+2) and to medium (-1,+1) positions.

LOD Bias = -2

LOD Bias = -1

LOD Bias = 0

LOD Bias = 1

LOD Bias = 2

As you can see, the best is something about -0.5. But you'll have to do such operations for every other game and that's not good. There's no unique LOD Bias for all games, so you'll have to search for it in every one of them. In addition, it's hard to get a good detail level along with good scene geometry on Voodoo4-5. Here's GeForce2 with LOD Bias set by default:

Above is the Unreal (Direct3D) screenshot, taken with NVIDIA GeForce2 GTS-based videocard with trilinear filtering enabled. Below is the same but with Glide on 3dfx Voodoo5 5500 with default settings:

LOD Bias is lowered and it spoils the floor. It changes to better only with LOD Bias +2, but this spoils walls. Anyway it lacks quality:

3dfx Voodoo5 5500 is to the left, NVIDIA GeForce2 GTS is to the right

But despite all this, picture quality of 3dfx Voodoo5 5500 is on a good level, though some new games show bugs.

Giants have just disappointed us.

Moreover, there are Depth Precision options for 16/32-bit color, they allow to correct precision during color depth calculations. This is set to 'Fast' by default, what spoils some calculations. Changing this feature gives a noticable speed boost:

But below is the price for such settings. Here are two screenshots taken with the highest/lowest Depth Precision settings (16-bit is to the left, 32-bit is to the right):

Disable:

Faster:

You can clearly see artefacts in 16-bit color with Depth Precision set to "Faster", they are almost absent in 32-bit mode.

I want to note that there's a hidden feature in 1.04.01 drivers (and newer) concerning HSR (Hidden Surface Remove) technology that deletes some hidden surfaces in the game to free the GPU, but as there's no hardware support for this feature, this is done by the drivers and CPU.

This feature seems to work (unlike on Voodoo3), but how? Here are some diagrams of perfomance boost in percentage with HSR on at various levels:

See that boost? It rules. But don't hurry to switch this feature on. Let's look at quality with HSR enabled:

Aggressive Tiling: Semi-Aggressive Tiling:

Non-Aggressive Tiling: Conservative Tiling:

I shall say that only Conservative level is good enough, higher levels spoil the picture. But even 27% boost is very good. Still HSR works only with OpenGL, i.e. Unreal has no reaction for it and Direct3D has no such feature at all.

Good news: FSAA quality in Need For Speed: Porshe 2000 is outstanding along with nice gameplay:

In Mercedes Benz Truck Racing FSAA is also perfect:

A pity, but FSAA 4x greatly reduces gameplay in other games, and FSAA 2x is not considered as it's worse than NVIDIA's supersampling.

2D quality is just perfect, and it stands on NVIDIA GeForce2 level. You can observe a very clear picture in 1600x1200 resolution.

Write a comment below. No registration needed!

Article navigation:

blog comments powered by Disqus