RayN: Ray Tracing Acceleration with Near-memory Computing

A desire for greater realism and increasing transistor density has led the GPU industry to include specialized hardware for accelerating ray tracing in graphics processing units (GPUs). Ray tracing generates realistic images, but even with specialized hardware support for memory traversal, we find it is highly sensitive to memory latency due to pointer chasing operations. Noting that graphics has historically employed specialized DRAM memory (GDDR) and near-memory computing is highly suitable for improving memory access latency, we propose RayN, a near-memory ray tracing architecture. RayN places a memory controller and dedicated ray tracing computation units inside the logic layers of 3D stacked DRAM memory modules specialized for use with GPUs. We study different memory controller configurations for RayN and how to partition scene geometry among memory modules to mitigate load imbalance. RayN achieves 3.0 × speedup on average on representative ray tracing workloads over a GPU without near-memory support while incurring a minimal area and power overhead.