Powerful and reliable programming model and computing toolkit

Report Share

NVIDIA CUDA Toolkit 13.3.1 (for Windows 10)

August, 1st 2026 - 2.4 GB - Freeware

Free Download

Safe & Secure

Latest Version

NVIDIA CUDA Toolkit 13.3.1 (for Windows 11)
Operating System

Windows 10
User Rating

Click to vote
Author / Product

NVIDIA Corporation / External Link
Filename

cuda_13.3.1_windows.exe

Sometimes latest versions of the software can cause issues when installed on older devices or devices running an older version of the operating system.

Software makers usually fix these issues but it can take them some time. What you can do in the meantime is to download and install an older version of NVIDIA CUDA Toolkit 13.3.1 (for Windows 10).

For those interested in downloading the most recent release of NVIDIA CUDA Toolkit or reading our review, simply click here.

All old versions distributed on our website are completely virus-free and available for download at no cost.

We would love to hear from you

If you have any questions or ideas that you want to share with us - head over to our Contact page and let us know. We value your feedback!

Download NVIDIA CUDA Toolkit 13.3.1 (for Windows 10)

Why is this app published on FileHorse? (More info)

What's new in this version:

New Features:
- The TMA-based kernel (Hopper and newer) now accelerates DSYMV in addition to the already-enabled SSYMV. The 16-byte alignment requirement for the A pointer was dropped for this kernel, and support for atomics was added through cublasSetAtomicsMode(). The geomean speedup across architectures and datatypes is 1.3x, and up to 5.9x.

Fixed:
- Fixed an issue where cublasXt<t>spmm() could produce incorrect results with m greater than 46340. [6155165]
- Fixed an issue where cublasLtMatmul() could run an unsupported combination of data types: an FP32-like compute type with FP32 C and D and non-FP32 A and B, in which case A and B are incorrectly interpreted as FP32 matrices.
- Fixed an issue where cublasLtMatmul() returned CUBLAS_STATUS_NOT_SUPPORTED for FP8 Grouped GEMM problems with scale modes CUBLASLT_MATMUL_MATRIX_SCALE_VEC128_32F and CUBLASLT_MATMUL_MATRIX_SCALE_BLK128x128_32F on Hopper GPUs.
- Fixed an issue where cublasLtMatmul() with int8 inputs and scale type CUDA_R_32I would allow non-default epilogues on Blackwell sm10x GPUs and return incorrect results. The correct behavior is to disallow all but the default epilogue, as documented.