cuBLASDx

Type: Technology Tags: NVIDIA, CUDA, cuBLAS, BLAS, device library, linear algebra, CUDA-X Related: cuBLAS, NVIDIA-CUDA, CUTLASS, cuFFTDx, CUDA-Cpp-Standard-Library, nvmath-python Sources: https://docs.nvidia.com/cuda/cublasdx/ Last Updated: 2026-04-29

Summary

cuBLASDx is NVIDIA’s device-side BLAS-style library for integrating dense linear algebra operations into CUDA kernels. It complements cuBLAS, which provides host-launched BLAS APIs for GPU-accelerated dense linear algebra.

Detail

Purpose

Applications sometimes need BLAS-like operations fused with custom CUDA computation to avoid extra launches or memory round trips. cuBLASDx provides device-callable building blocks for those expert CUDA workflows.

Key capabilities

• Device-side dense linear algebra operations for CUDA C++ kernels.
• Documentation includes getting started material, API reference, Python information, examples, and references.
• Useful when developers want library-grade math blocks inside custom kernels.
• Adjacent to CUTLASS for custom GEMM construction and cuBLAS for full library APIs.

NVIDIA context

cuBLASDx reflects NVIDIA’s move toward composable device libraries for high-performance fused kernels, especially relevant to AI, HPC, and simulation workloads that mix standard math operations with custom code.

Connections

• cuBLAS - host-level BLAS library complemented by cuBLASDx.
• CUTLASS - template library for building custom matrix-multiply kernels.
• cuFFTDx - sibling device-side math library for FFT operations.
• nvmath-python - Python-facing NVIDIA math package can expose CUDA-X math capabilities at a higher level.
• NVIDIA-CUDA - cuBLASDx appears in CUDA math-library documentation.

Source Excerpts

• NVIDIA’s cuBLASDx docs provide getting started, API reference, Python, examples, and references sections.