cuBLASXt

Type: Technology Tags: NVIDIA, CUDA, cuBLAS, multi-GPU, BLAS3, linear algebra, CUDA-X Related: cuBLAS, cuBLASLt, cuBLASMp, NCCL, NVSHMEM, GPUDirect-RDMA, NVIDIA-HPC-SDK, NVIDIA-CUDA Sources: https://docs.nvidia.com/cuda/cublas/index.html#using-the-cublasxt-api Last Updated: 2026-04-29

Summary

cuBLASXt is NVIDIA’s cuBLAS host interface for single-node, multi-GPU BLAS Level 3 operations. It lets applications provide matrices in host memory or participating device memory, select a set of GPUs, and have the library distribute compute-intensive matrix operations across those GPUs.

Detail

Purpose

Large dense matrix operations can exceed the practical capacity or throughput of one GPU, but rewriting applications around explicit multi-GPU tiling is expensive. cuBLASXt provides a higher-level host API that manages multi-GPU allocation, transfers, tiling, scheduling, and result collection for supported BLAS3 routines.

Key capabilities

• Multi-GPU host interface exposed through cublasXt.h.
• Supported for compute-intensive BLAS Level 3 and BLAS-like routines where transfer overhead can be amortized.
• User-selected GPU set through cuBLASXt context and device-selection APIs.
• Tiling strategy with user-controllable block dimension and round-robin assignment across GPUs.
• One CPU thread per selected GPU to coordinate transfers and cuBLAS work for that device.
• Optional hybrid CPU-GPU execution controls for supported routines.
• 64-bit platform support requirement in the current NVIDIA documentation.

NVIDIA context

cuBLASXt occupies the single-node multi-GPU layer of NVIDIA dense linear algebra. It is adjacent to cuBLASLt for highly tuned GEMM on a library API, cuBLASMp for distributed multi-process dense linear algebra, and cluster communication layers such as NCCL and NVSHMEM when workloads move beyond one host.

Connections

• cuBLAS - cuBLASXt is one of the API families exposed by the cuBLAS library.
• cuBLASLt - sibling API for programmable GEMM and algorithm heuristics.
• cuBLASMp - multi-process distributed dense linear algebra path beyond cuBLASXt’s single-node model.
• NCCL - collective communication layer commonly used in adjacent multi-GPU workloads.
• NVSHMEM - GPU-cluster programming model used in scale-out NVIDIA math libraries.
• NVIDIA-HPC-SDK - packages CUDA math libraries for HPC development and deployment workflows.

Source Excerpts

• NVIDIA documents cuBLASXt as a multi-GPU capable cuBLAS host interface that dispatches BLAS3 work across selected GPUs.
• The current cuBLAS docs describe cuBLASXt tiling, device selection, CPU-GPU hybrid controls, and supported BLAS3 math functions.