NVIDIA CUDA

The programming support for NVIDIA GPUs in Julia is provided by theCUDA.jl package. It is built on the CUDA toolkit, and aims to be as full-featured and offer the same performance as CUDA C. The toolchain is mature, has been under development since 2014 and can easily be installed on any current version of Julia using the integrated package manager.

CUDA.jl makes it possible to program NVIDIA GPUs at different abstraction levels:

• by using theCuArray type, providing a user-friendly yet powerful abstraction that does not require any GPU programming experience;

• by writing CUDA kernels, with the same performance as kernels written in CUDA C;

• by interfacing with CUDA APIs and libraries directly, offering the same level of flexibility you would expect from a C-based programming environment.

Thedocumentation of CUDA.jl demonstrates each of these approaches.

Performance

Julia on the CPU is known for its good performance, approaching that of statically compiled languages like C. The same holds for programming NVIDIA GPUs with kernels written using CUDA.jl, where we haveshown the performance to approach and even sometimes exceed that of CUDA C on a selection^[1] of applications from the Rodinia benchmark suite:

Relative performance of Rodinia benchmarksimplemented in Julia with CUDA.jl.