Support Matrix

GPU, CUDA Toolkit, and CUDA Driver Requirements

The following sections highlight the compatibility of NVIDIA cuDNN versions with the various supported NVIDIA CUDA Toolkit, CUDA driver, and NVIDIA hardware versions.

Supported NVIDIA Hardware and CUDA Version

cuDNN Package 1

CUDA Toolkit Version

Supports static linking? 2

NVIDIA Driver Version for Linux

NVIDIA Driver Version for Windows

CUDA Compute Capability

Supported NVIDIA Hardware

cuDNN 9.0.0 for CUDA 12.x

12.3

Yes

>=525.60.13

>=527.41

  • 9.0 3

  • 8.9 3

  • 8.6

  • 8.0

  • 7.5

  • 7.0

  • 6.1

  • 6.0

  • 5.0

  • NVIDIA Hopper 3

  • NVIDIA Ada Lovelace architecture 3

  • NVIDIA Ampere architecture

  • NVIDIA Turing

  • NVIDIA Volta

  • NVIDIA Pascal

  • NVIDIA Maxwell

cuDNN 9.0.0 for CUDA 12.x

  • 12.2

  • 12.1

  • 12.0

No

>=525.60.13

>=527.41

  • 9.0 3

  • 8.9 3

  • 8.6

  • 8.0

  • 7.5

  • 7.0

  • 6.1

  • 6.0

  • 5.0

  • NVIDIA Hopper 3

  • NVIDIA Ada Lovelace architecture 3

  • NVIDIA Ampere architecture

  • NVIDIA Turing

  • NVIDIA Volta

  • NVIDIA Pascal

  • NVIDIA Maxwell

cuDNN 9.0.0 for CUDA 11.x

11.8

Yes

>= 450.80.02

>=452.39

  • 9.0 3

  • 8.9 3

  • 8.6

  • 8.0

  • 7.5

  • 7.0

  • 6.1

  • 6.0

  • 5.0

  • NVIDIA Hopper 3

  • NVIDIA Ada Lovelace architecture 3

  • NVIDIA Ampere architecture

  • NVIDIA Turing

  • NVIDIA Volta

  • NVIDIA Pascal

  • NVIDIA Maxwell

cuDNN 9.0.0 for CUDA 11.x

  • 11.7

  • 11.6

  • 11.5

  • 11.4

  • 11.3

  • 11.2 4

  • 11.1 4

  • 11.0 4

No

>= 450.80.02

>=452.39

  • 9.0

  • 8.9

  • 8.6

  • 8.0

  • 7.5

  • 7.0

  • 6.1

  • 6.0

  • 5.0

  • NVIDIA Hopper

  • NVIDIA Ada Lovelace architecture

  • NVIDIA Ampere architecture

  • NVIDIA Turing

  • NVIDIA Volta

  • NVIDIA Pascal

  • NVIDIA Maxwell

Note

For best performance, the recommended configuration for GPUs Volta or later is cuDNN 9.0.0 with CUDA 12.3. For GPUs prior to Volta (that is, Pascal and Maxwell), the recommended configuration is cuDNN 9.0.0 with CUDA 11.8. These are the configurations used for tuning heuristics.

CPU Architecture and OS Requirements

The following tables highlight the compatibility of cuDNN versions with the various supported OS versions.

Linux

Refer to the following table to view the list of supported Linux versions for cuDNN.

Linux Versions for cuDNN

Architecture

OS Name

OS Version

Distro Information: Kernel

Distro Information: GCC

Distro Information: Glibc

x86_64

RHEL

9.x

5.14.0

8.5.0

2.34

x86_64

RHEL

8.x

4.18.0

8.5.0

2.28

x86_64

RHEL

7.x

3.10.0

4.8.5

2.19

x86_64

Rocky

8.6

4.18

8.5.0

2.28

x86_64

Ubuntu

22.04

5.15.0

11.2.0

2.35

x86_64

Ubuntu

20.04

5.15.0

9.4.0

2.31

x86_64

Debian

11.4

5.10

10.2.1

2.31

AArch64 incorporates ARM based CPU cores for Server Base System Architecture (SBSA).

RHEL

9.x

5.14.0

8.5.0

2.23

AArch64 incorporates ARM based CPU cores for Server Base System Architecture (SBSA).

RHEL

8.x

4.18

8.5.0

2.28

AArch64 incorporates ARM based CPU cores for Server Base System Architecture (SBSA).

Ubuntu

22.04

5.15.0

11.2.0

2.35

AArch64 incorporates ARM based CPU cores for Server Base System Architecture (SBSA).

Ubuntu

20.04

5.14.0

9.3.0

2.31

Note

  • For platforms that ship a compiler version older than GCC 6 by default, linking to static cuDNN using the default compiler is not supported.

  • For RHEL 8.9 and Rocky 8.9 Linux, the R525 and later display driver is needed.

Windows

Windows 10 and Windows Server 2019 and 2016 are supported. Refer to the following table to view the list of supported Visual Studio versions for cuDNN.

Visual Studio Versions Based on Your Version of CUDA

CUDA 12.x - 11.8

CUDA 11.7 - 11.0

Visual Studio

2019

2017

Footnotes

1

For the dynamic cuDNN libraries, the cuDNN build for CUDA 12.x is compatible with CUDA 12.x for all x, including future CUDA 12.x releases that ship after this cuDNN release. Similarly, the cuDNN build for CUDA 11.x is compatible with CUDA 11.x for all x. For the limitation when using the static cuDNN library, refer to this table and the cuDNN Release Notes for more information.

2

This column specifies whether the given cuDNN library can be statically linked against the CUDA toolkit for the given CUDA version. Dynamic linking is supported in all cases.

3(1,2,3,4,5,6,7,8,9,10,11,12)

Requires CUDA Toolkit >= 11.8.

4(1,2,3)

Functionalities that perform runtime compilation such as the runtime fusion engines, and the persistent dynamic algorithm of RNN, requires NVRTC from CUDA Toolkit 11.2 update 1 or later.