Multi-GPU multi-display rendering of extremely large 3D environments

In real-time rendering applications, mesh rendering quality suffers from limited GPU memory capacity and display resolution. Due to the increased complexity of models and the demand for higher display resolutions, people have started building commodity workstations with multiple GPUs at a low cost. As a result, more GPU memory is available across multiple GPUs, and a higher display resolution can be achieved by connecting each GPU to a display monitor, resulting in a large tiled display configuration. However, a multi-GPU workstation may not efficiently handle a complex model that cannot fit into the GPU memory, due to (1) the unified configuration treating GPUs as one hardware entity and requiring the same data replicated in all GPUs, and (2) the lack of scalability to reduce, balance, and stream data dynamically between the CPU and GPUs as well as among the GPUs. In this work, we present a fine-grained parallel rendering approach that integrates a view-dependent LOD selection strategy with the inter-GPU load balancing method to ensure each GPU handles the portion of data it rasterizes, without data replication. A new multi-GPU out-of-core method minimizes the amount of data transferred from the CPU to each GPU by taking the advantage of frame-to-frame coherence. A comprehensive evaluation is presented to understand the efficiency and scalability of the execution components over extremely large scenes.