cuda

CUDA stands for Compute Unified Device Architecture. It is a parallel computing platform and programming model developed by NVIDIA that enables developers to use NVIDIA GPUs for general-purpose computing (GPGPU). CUDA provides a software abstraction for writing kernels that run on the GPU and manage data transfers between host memory and device memory.

The CUDA platform includes a compiler, libraries, and tools as part of the CUDA Toolkit. Kernels are written in languages such as CUDA C/C++ (and, in some cases, CUDA Fortran), compiled with the nvcc compiler, and launched from host code. Programs expose a host-device execution model where the CPU coordinates work for many parallel GPU threads.

Execution is organized into grids of thread blocks; each thread executes a kernel on a streaming multiprocessor of the GPU. CUDA exposes a hierarchy of memory types, including global, shared, constant, and texture memory, with mechanisms for synchronization and asynchronous execution through streams and events. Unified Memory simplifies memory management by allowing the system to migrate data between host and device as needed.

NVIDIA provides a set of libraries optimized for CUDA, such as cuBLAS for linear algebra, cuFFT for fast Fourier transforms, cuDNN for deep neural networks, and NCCL for multi-GPU communication. The CUDA Toolkit also includes profiling and debugging tools like Nsight and nvprof.

CUDA is widely used for high-performance computing, scientific simulations, and accelerated machine learning workloads. Adoption is generally tied to NVIDIA GPUs, and portable alternatives exist (e.g., OpenCL, SYCL). The platform continues to evolve with new GPU architectures and expanded tooling.