Now we've seen what these methods all look like but we haven't really looked at what kind of performance implications they have. Unfortunately in our comparisons below we couldn't include Anisotropic filtering as no hardware really supports it, the TNT's drivers reports Anisotropic filtering capabilities but when rendering it does only tri-linear.

3dfx Voodoo 2
Point Sample 100.8 - 101.2% Bi-linear 100% Tri-linear 65.9 - 73.1%

Matrox G200
Point Sample 106.4 - 111.4% Bi-linear 100% Tri-linear 68.7 - 89.0%

nVidia RIVA TNT
Point Sample 100.9 - 101.4% Bi-linear 100% Tri-linear 81.8 - 92.3%

S3 Savage 3D
Point Sample 102.0 - 102.9% Bi-linear 100% Tri-linear 100.0 - 100.5%

As the above benchmarks show using Point sampling usually gives a slightly performance boost but the visual quality suffers. Bi-linear filtering performance represents the 100% mark. Tri-linear filtering is up to 35% slower on some solutions but less on others. The Savage3D as you can see does 100% with tri-linear due to its single pass single cycle tri-linear filtering feature.

One thing that all accelerators above hold in common is that the performance impact when using tri-linear filtering is higher the higher the resolution you use hence the lowest scores have been achieved at 1600 x 1200 for the most part (on the Voodoo2 at 800 x 600 as it doesn't support higher resolutions in non SLI mode).

That's it for this time, questions, comments? Send them to Anders Hammervald, Until next time...