next up previous
Next: Conclusions Up: Graphics Subsystem Bottlenecks Previous: OpenGL Command Bandwidth

Beware of Benchmarks

Beware of graphics benchmarks. The performance that matters is the performance of your applications. Graphics benchmarks are often difficult to interpret because not enough parameters are specified. For example, a vendor may claim ``10 million 3D lines per second.'' While this number is high, it does not say if the lines are depth buffered, if the lines are connected or unconnected,gif if the lines are shaded or textured, etc. Also, if your application is rendering polygons, line rendering speed is probably irrelevant.

Graphics benchmarks often rely on synthetic workloads. Real graphics programs do not simply render millions of 3D lines. Instead, real programs switch between various primitives and use various modes, and this switching overhead is rarely measured by benchmarks. Real programs also spend time generating the data to render, a task not done in benchmarks. Benchmarks are useful for determining the ``not to be exceeded" rendering performance for a particular rendering task, but do not expect real applications to achieve benchmark results.



mjk@sgi.com