The GPU engine module runs array computations on the GPU through WebGPU (Dawn), portably on Windows, Linux and macOS without a proprietary CUDA dependency.

Arrays are moved to the device with gpuArray and back to the host with gather. Operators and many functions are overloaded so that expressions on a gpuArray run on the device and keep their result device-resident.

Device arrays are stored in single precision (WebGPU has no double compute type yet); real, logical and complex data are supported. Use canUseGPU to check availability and gpuDevice to inspect the selected device.

Many functions are supported on a gpuArray: the operators, the element-wise math functions (including the rounding and hyperbolic/inverse-trigonometric ones), the reductions and scans (sum, prod, mean, var, std, cumsum, cumprod, sort, find), the complex-extraction functions, the shape and predicate functions, and the constructors zeros/ones/rand/randn through the 'gpuArray' / 'like' forms. Call gpuArrayFunctions for the exact, always-current list.