LOSTCIRCUITS

Command Processor

One of the centerpieces of the Wildcat VP architcture is the 3DLabs multithreaded Command Processor. Compared to more conventional architectures, this is the first command processor to bring multi-threading capabilities to a graphics processor by monitoring command buffers from multiple threads and applications on the host CPU. If one buffer runs dry the command processor will autonomously find a buffer that contains useful work, without interrupting the CPU. Normally a graphics device needs to interrupt the CPU to determine what to do when a buffer runs empty. This can overload the CPU in case it is not keeping up in the first place. Command buffer processing allows the Wildcat VP technology to efficiently multi-task by switching and servicing the host applications that require it. As a consequence, the Wildcat VP technology requires less CPU cycles.

Data, including multithreaded commands reach the graphics adapter via the AGP interface, are loaded into the local frame buffer (video memory) and then the command processor.

Vertex Processor

Most graphics processors are using four element vector processors, that is four individual processing units are hardwired together since most graphics entities consist of four individual elements. An alternative solution to provide more flexibility at the expense of a minor performance hit is to keep each of the scalar processors separate. In turn, this allows to work more efficiently on scalar or two and three element arrays that would run at 25, 50 or 75% efficiency on a fixed four element vector processor. By the end of the day, though, it is redundancy vs. efficiency, meaning that the one size fits all approach may be universally applicable but may have some other problems.

Sixteen scalar vertex processing units offer more flexibility than four ganged vector4 units, however, the arrangement is completely transparent (invisible) to the controller.

In theory, whether there are four 4-element vector processors or 16 scalar units should come out with a net-zero difference since four element vectors get processed by four independent scalar processors as quickly as if a single four element vector processor were used. Moreover, either mode of parallelism is completely transparent to the compiler, that is the compiler effectively does not know and neither does it care what is inside the black box.between the vertex processor manager (to distribute the workload amongst the different scalar / vector units) and the vertex processor multiplexer (VPMux) to recombine the data. In practice, a 4-element vector processor has less overhead than four single scalar processsors.

Next Page:    => Clipping, Culling => Rasterization and Texturing =>