Repurposing Hardware Ray-Tracing Accelerators for GPGPU Workloads

Modern GPUs include specialized fixed-function application-specific hardware units. Two prominent examples in recent NVIDIA architectures are:

Tensor Cores: High-density matrix multiply-accumulate (MMA) systolic engines.
RT (Ray-Tracing) Cores: Dedicated hardware for bounding volume hierarchy (BVH) traversal and ray-triangle intersection testing.

While Tensor Cores have been heavily explored for scientific computing and linear algebra outside deep learning, RT Cores remain largely under-utilized for non-graphics computing. In our current research in the ESCAL Lab, we are investigating how to map irregular tree and spatial query workloads directly onto RT Core hardware via NVIDIA OptiX.

Why RT Cores?

Ray tracing fundamentally consists of two operations:

Hierarchical BVH Traversal: Finding which bounding boxes in 3D space intersect a ray.
Primitive Intersection: Testing whether a ray intersects a primitive (e.g. triangle, custom geometry).

In modern GPU silicon, BVH tree traversal is offloaded from general-purpose Streaming Multiprocessors (SMs) directly to hardware traversal units within the RT cores. This frees SM execution pipelines and ALU registers.

Target Non-Graphics Workloads

Gradient Boosted Decision Trees (GBDT): Inference in decision tree ensembles (e.g., XGBoost, LightGBM) requires navigating millions of binary decision trees. By encoding the tree structure into 3D bounding boxes and query vectors into rays, tree evaluation can be framed as a BVH traversal query.
Multiple Sequence Alignment (MSA): Aligning biological sequences often maps to search spaces over multi-dimensional grids. Hardware-accelerated ray intersection tests can be used to prune suboptimal alignments quickly.
Spatial Indexing & Nearest Neighbor Searches: k-d tree queries and range searches can be naturally represented as ray-box and bounding volume checks.

Stay tuned for more updates as our prototypes progress!