NVIDIA GeForce3 Investigation: What NVIDIA didn't tell us
by Kert Chian on May 15, 2001 1:18 AM EST- Posted in
- GPUs
Z-compression, Z-occlusion culling and Crossbar memory controller
By implementing 4:1 lossless data compression, Z-buffer bandwidth is reduced by a factor of four. In the 3D Mark 2000 fillrate test, translucent quadrilaterals that occupy the entire screen are rendered repeatedly. Because the triangles are translucent and therefore do not occlude one another, Z-occlusion culling does not interfere with fillrate measurement. In addition, framebuffer bandwidth is well-utilized because the test consists of large triangles that occupy half the screen resolution. By this process of exclusion, the vast 32-bit fillrate improvements as is obvious in figure 1 may be fully attributed to Z-compression. The only surprise is 16-bit fillrate, which trails GeForce2 slightly. The likely explanation is that there are no bandwidth constraints in the 16-bit test, and that GeForce2 has better fillrate efficiency (at least in 16-bit mode).
Figure 1: 3D Mark 2000 multi-texture fillrate for GeForce2 and GeForce3
16-bit fillrates | 32-bit fillrates | |||
Resolution | GeForce2 | GeForce3 | GeForce2 | GeForce3 |
640x480 |
1293.2
|
1212.3
|
743.7
|
1176.2
|
800x600 |
1335.8
|
1212.9
|
750.7
|
1176.6
|
1024x768 |
1388.5
|
1278
|
782.8
|
1244.8
|
1280x1024 |
1405.2
|
1345.5
|
739.4
|
1292.2
|
1600x1200 |
1408.1
|
1391.5
|
717.5
|
1327.5
|
The vastly superior 32-bit fillrate, as measured in 3D Mark 2000, is in
fact borne out in real-world benchmarks. As a matter of fact, GeForce3 32-bit
fillrate frequently exceeds GeForce2 16-bit fillrate! The high polygon counts
and wasted overdraw of such benchmarks provide ample opportunities for further
optimizations, and serve as indirect proof of the existence of Z-occlusion
culling and crossbar memory controllers.
0 Comments
View All Comments