OpenGL Insights

Fermi Asynchronous

Texture Transfers

Shalini Venkataraman

29.1 Introduction

Many real-world graphics applications need to transfer textures efficiently in and

out of the GPU memory in the form of 2D images, 2.5D terrains, or 3D volumes

and their time-varying counterparts. In the pre-Fermi generation of NVIDIA hard-

ware, any data transfer would stall the GPU from rendering because the GPU had

a single hardware execution thread that could either execute transfers or rendering.

The OpenGL pixel buffer object (PBO) [ARB 04] provides a mechanism to optimize

transfers, but it is CPU asynchronous in that it allows for concurrent CPU processing

while the GPU