This paper presents a fast and efficient method for ray tracing particle-based scene representations such as 3D Gaussian Splatting. The key idea is to construct encapsulating primitives around each particle and insert them into a bounding volume hierarchy (BVH) to be rendered by a ray tracer optimized for high-density overlapping particles. This approach enables efficient ray tracing, which opens the door to advanced techniques like secondary ray effects (mirrors, refractions, shadows), highly-distorted cameras with rolling shutter effects, and stochastic sampling of rays. The proposed method is implemented using NVIDIA OptiX and is designed to be efficient and differentiable, allowing for optimization of particle scenes from observed data. The method involves using adaptive bounding mesh primitives to leverage fast ray-triangle intersections and shading batches of intersections in depth-order. The algorithm is tested on various benchmarks and applications, demonstrating speed and accuracy. The paper also proposes improvements to the basic Gaussian representation, including the use of generalized kernel functions that significantly reduce particle hit counts. The paper evaluates the proposed approach on a wide variety of benchmarks and applications, showing that ray tracing nearly matches or exceeds the quality of the 3DGS rasterizer while achieving real-time rendering framerates. It also demonstrates new techniques made possible by ray tracing, including secondary ray effects, rendering from highly-distorted cameras, and training with stochastically sampled rays. The paper compares the proposed method with existing approaches, including neural radiance fields (NeRFs), point-based and particle rasterization, and differentiable ray tracing of volumetric particles. It highlights the limitations of rasterization, such as the inability to represent highly distorted cameras, model secondary lighting effects, or simulate sensor properties like rolling shutter or motion blur. The proposed method addresses these limitations by using optimized ray tracing throughout both training and inference, allowing for complex effects like depth of field and perfect mirrors. The paper also discusses the use of different particle kernel functions, including generalized Gaussians, kernelized surfaces, and cosine wave modulations, which can reduce the number of intersections and improve rendering efficiency. The results show that the proposed method outperforms existing approaches in terms of speed and quality, particularly for densely-clustered multi-view scenes. The method is evaluated on various benchmarks, including MipNeRF360, Tanks & Temples, Deep Blending, and NeRF Synthetic, demonstrating its effectiveness in novel-view synthesis and rendering. The paper concludes that the proposed method provides a key algorithmic ingredient for future research on particle-based scene representations.This paper presents a fast and efficient method for ray tracing particle-based scene representations such as 3D Gaussian Splatting. The key idea is to construct encapsulating primitives around each particle and insert them into a bounding volume hierarchy (BVH) to be rendered by a ray tracer optimized for high-density overlapping particles. This approach enables efficient ray tracing, which opens the door to advanced techniques like secondary ray effects (mirrors, refractions, shadows), highly-distorted cameras with rolling shutter effects, and stochastic sampling of rays. The proposed method is implemented using NVIDIA OptiX and is designed to be efficient and differentiable, allowing for optimization of particle scenes from observed data. The method involves using adaptive bounding mesh primitives to leverage fast ray-triangle intersections and shading batches of intersections in depth-order. The algorithm is tested on various benchmarks and applications, demonstrating speed and accuracy. The paper also proposes improvements to the basic Gaussian representation, including the use of generalized kernel functions that significantly reduce particle hit counts. The paper evaluates the proposed approach on a wide variety of benchmarks and applications, showing that ray tracing nearly matches or exceeds the quality of the 3DGS rasterizer while achieving real-time rendering framerates. It also demonstrates new techniques made possible by ray tracing, including secondary ray effects, rendering from highly-distorted cameras, and training with stochastically sampled rays. The paper compares the proposed method with existing approaches, including neural radiance fields (NeRFs), point-based and particle rasterization, and differentiable ray tracing of volumetric particles. It highlights the limitations of rasterization, such as the inability to represent highly distorted cameras, model secondary lighting effects, or simulate sensor properties like rolling shutter or motion blur. The proposed method addresses these limitations by using optimized ray tracing throughout both training and inference, allowing for complex effects like depth of field and perfect mirrors. The paper also discusses the use of different particle kernel functions, including generalized Gaussians, kernelized surfaces, and cosine wave modulations, which can reduce the number of intersections and improve rendering efficiency. The results show that the proposed method outperforms existing approaches in terms of speed and quality, particularly for densely-clustered multi-view scenes. The method is evaluated on various benchmarks, including MipNeRF360, Tanks & Temples, Deep Blending, and NeRF Synthetic, demonstrating its effectiveness in novel-view synthesis and rendering. The paper concludes that the proposed method provides a key algorithmic ingredient for future research on particle-based scene representations.