Vortex : deferred sort last parallel graphics architecture

  • Abram Santilli

Western Sydney University thesis: Doctoral thesis

Abstract

We have developed a new cluster parallel graphics architecture that improves upon prior cluster parallel graphics systems for high performance supergraphics - the Vortex deferred sort-last parallel graphics architecture based on global Zspace paradigm. The new architecture bypasses limitations of screen-space based parallelization paradigms, solves known Z-space parallelization inefficiencies and problems. Vortex addresses the lack of global Z buffer awareness between GPUs and prevents artifacts in globally order-dependent blending on multiple GPUs. The new paradigm allows for full 1-1 process-GPU coupling with minimal interprocess and inter-GPU communications. This allows for maximal input bandwidth, maximal GPU utilization levels, near optimal load balances, and improved efficiency when scaled to larger configurations. The Vortex architecture introduces the new deferred sort-blend approach for preventing visual artefacts in globally order-dependent fragment blends. All blend fragments are buffered in an external sort-blend subsystem until the end of rendering, when they are Z-culled, sorted and blended into the final frame. The new approach allows for efficient automatic order-independent blending and results in frames without any global blending artifacts The new architecture gives us the ability to fully harness the processing power of state of art GPUs, and at the same time it offers a much easier parallelization paradigm to programmers, compared to existing screen-space CPGS parallelization paradigms.
Date of Award2006
Original languageEnglish

Keywords

  • computer graphics
  • image processing
  • real-time data processing

Cite this

'