Dieter Schmalstieg
Is It Time to Move Beyond the Rasterization Pipeline?
Since its inception, the classic rasterization pipeline on the GPU has seen many extension in the form of new shaders, but it remains a hybrid of programmable and fixed function stages. New shader types and capabilities are added as new requirements emerge, such as low-power mobile operation, neural inference or raytracing. However, these novelties tend to be added locally (e.g., as a new pipeline stage) rather than as a globally available feature. Alas, combining the desired features as concatenated render passes is not very efficient. Meanwhile, entire graphics pipelines, such as Unreal's Nanite, are implemented entirely in compute shaders. This talk will review important directions in alternative pipeline stages and propose a potential direction where all stages of the pipeline are programmable, potentially unlocking dramatically better throughput. This idea will be illustrated by some selected studies from recent research.
back to overview