Fondecyt Regular 1221357
Combining Tensor Cores with Ray-Tracing Cores for GPU-based Simulations
ABSTRACT
The main objective of this proposal is to develop new approaches that can improve the performance and energy-efficiency of GPU-based scientific simulations, by combining tensor-core with RT-core operations. The hypothesis is that a significant part of the work involved in GPU-based scientific simulations can be mapped as fast MMA (tensor-core) and space-search operations in a BVH data structure (RT-core), improving their performance and energy-efficiency. Furthermore, the combination of these cores can produce new energy-efficient computational patterns. The specific objectives consist of redesigning frequently used GPU computational patterns such as nearest-neighbors exploration in structured-space, nearest-neighbors search in unstructured- space, sorting/min/max, elementary graph operations and to compose a new computational pattern for semi- structured space. The methodology is based on representing local/short-range computations as structured- space tensor-core MMA operations, codifying work into the row-column computations, and global/long-range computations as unstructured-space RT-core ray-triangle intersections. In addition to the computational patterns already mentioned in the specific objectives, the combination of these two approaches would allow the formulation of a new computational pattern for semi-structured-space simulations. The expected results are a major improvement in performance and energy-efficiency for the computational patterns, allowing the study of larger simulations as well as enable real-time performance for some of them. Numerical precision will be studied as well, finding how much does the problem size, data types and data distributions affect the result. The importance of this research is that the improvements produced on each computational pattern affects not only one specific scientific simulation, but families of simulations from the different fields of science.