Based on the TREEMARK tests, fifty to one hundred thousand polygons per frame can be expected to run at reasonable framerates (30 to 60 fps) on a 450 MHz Celeron. This is the case with one hardware light (directed, point
or spot). The effect of multiple hardware lights on polygon throughput is covered later on. There are, of course, other bottlenecks that can affect polygon throughput like bandwidth, fillrate and engine efficiency.
To realize the advantages of hardware T & L, static geometry is stored in local video memory. This conserves AGP bandwidth and offloads T & L computations from the CPU, allowing the reclaimed CPU horsepower to be used for things like improved physics, artificial intelligence and character animation.
By moving static geometry from AGP memory to local video memory, local memory bandwidth is now taxed with an additional geometry data stream, in addition to framebuffer and texture datastreams. An estimate of this geometry data bandwidth would be in the region of two hundred to five hundred megabytes per second, at a polygon throughput of 1.5 to 3 million triangles per second.
|
Estimation of Vertex Data Bandwidth for a Direct3D Application
|
Static geometry database with 50,000 triangles (or 150,000 vertices).
This is stored in local video memory as Direct3D Vertex Buffers.
A Vertex Buffer containing 1000 vertices is approximately 48 to 64 kilobytes.
Therefore, 150 Vertex Buffers containing 1000 vertices are needed to store all 50,000 triangles, equivalent to 7200 to 9600 kilobytes of local video memory storage.
If the application runs at 30 to 60 frames per second, the bandwidth needed is 210 to 560 megabytes per second.
|
Character models are examples of dynamic geometry. Each frame of the animation is computed on the CPU. The animated vertices for each frame can be transformed by the CPU or sent through the AGP bus to be transformed by
the T & L hardware. Generally, using the CPU for transformation of dynamic geometry is more efficient than using hardware T & L. This is because the hardware will usually have to wait for the polygons to be uploaded across the AGP bus onto local video memory on a per frame basis. This will, in all likelihood, negate the performance advantage of T & L hardware.
GeForce has rudimentary support for vertex blending and is the first consumer graphics chip to do so. Vertex blending allows game developers to create smooth joints between the multiple parts of a geometric model. It is particularly useful for lifelike character animation.