cuBLAS

cuBLAS is a GPU-accelerated Basic Linear Algebra Subprograms (BLAS) library developed by NVIDIA as part of the CUDA Toolkit. It provides high-performance implementations of BLAS Level 1 through Level 3 routines for NVIDIA GPUs, enabling operations such as vector operations, matrix-vector products, and matrix-matrix multiplies on device memory. The library supports single and double precision, real and complex data types, and offers batched and strided batched variants to process multiple problems concurrently.

The API revolves around a cuBLAS handle, created with a function like cublasCreate and destroyed with cublasDestroy. Programs can associate a CUDA stream with the handle and choose a pointer mode (host or device). Routines are invoked with device pointers, and cuBLAS returns status codes of type cublasStatus_t for error handling. Common routine families include cublasSgemm, cublasDgemm, cublasCgemm, and cublasZgemm for general matrix-matrix multiplication, as well as cublasSgemv and cublasDgemv for matrix-vector multiplication. Higher-level operations are composed from these building blocks, and many routines provide batched variants to improve throughput on many small problems.

cuBLAS operates exclusively on GPU memory; data must reside in device memory and results are written back there. It is widely used in scientific computing, simulations, and machine learning workloads that require fast linear algebra. The library emphasizes compatibility with CUDA streams for asynchronous execution and performance tuned to NVIDIA architectures. NVIDIA also offers extensions such as cuBLASXt for multi-GPU workflows and extensive documentation across language bindings for C, C++, and Fortran.