CuDNN: Difference between revisions

From HPCWIKI
Jump to navigation Jump to search
No edit summary
 
(One intermediate revision by the same user not shown)
Line 3: Line 3:
One of that is [https://docs.nvidia.com/deeplearning/cudnn/pdf/cuDNN-Support-Matrix.pdf cuDNN support matrix] which should check before build CUDA related project code.
One of that is [https://docs.nvidia.com/deeplearning/cudnn/pdf/cuDNN-Support-Matrix.pdf cuDNN support matrix] which should check before build CUDA related project code.


== Linux Versions for cuDNN ==
== cuDNN version matrix ==
<big>*For best performance, the recommended configuration is</big>
 
<big>'''cuDNN 8.8.0 on H100 with CUDA 12.0, cuDNN 8.8.0 on all other GPUs with CUDA 11.8, because this is the configuration that was used for tuning heuristics.'''</big>
 
{| class="wikitable"
{| class="wikitable"
!cuDNN Version
!cuDNN Version
Line 16: Line 20:
!Glibc Version
!Glibc Version
|-
|-
| rowspan="3" |cuDNN 8.3.x
|cuDNN 8.8.1 for CUDA 12.x
| rowspan="12" |NVIDIA GPU with  
|NVIDIA Hopper™3
Compute Capability 5.0 or higher
|CUDA Toolkit 12.0
|
|
|
|
|
|
|
|-
|cuDNN 8.8.1 for CUDA 11.x
|NVIDIA Ada Lovelace Architecture4
NVIDIA Ampere Architecture
NVIDIA Turing™
NVIDIA Volta™
NVIDIA Pascal
NVIDIA GPU with Compute Capability SM 5.0 - SM 8.7
|CUDA Toolkit 11.8 or later
|
|
|
|
|
|
|
|-
| rowspan="4" |cuDNN 8.3.x
| rowspan="13" |NVIDIA GPU with  
Compute Capability 3.5 or higher
Compute Capability 3.5 or higher
| rowspan="3" |CUDA Toolkit 11.5 or later
| rowspan="4" |CUDA Toolkit 11.5 or later
| rowspan="3" |CUDA Driver 450.80.02 or later
| rowspan="4" |CUDA Driver 450.80.02 or later
| rowspan="16" |x86_64
| rowspan="17" |x86_64
|Ubuntu
|Ubuntu
|20.04 LTS
|20.04 LTS
Line 33: Line 65:
|7.5.0
|7.5.0
|2.27
|2.27
|-
|
|
|
|
|
|-
|-
|CentOS
|CentOS

Latest revision as of 16:20, 19 March 2023

NVIDIA products has lots of compatibility issues inbetween CUDA, cuDNN, driver and NVIDIA hardware by architecture.

One of that is cuDNN support matrix which should check before build CUDA related project code.

cuDNN version matrix

*For best performance, the recommended configuration is

cuDNN 8.8.0 on H100 with CUDA 12.0, cuDNN 8.8.0 on all other GPUs with CUDA 11.8, because this is the configuration that was used for tuning heuristics.

cuDNN Version GPU Requirements CUDA Toolkit Requirements CUDA Driver Requirements Architecture OS Name OS Version Kernel Version gcc Version Glibc Version
cuDNN 8.8.1 for CUDA 12.x NVIDIA Hopper™3

Compute Capability 5.0 or higher

CUDA Toolkit 12.0
cuDNN 8.8.1 for CUDA 11.x NVIDIA Ada Lovelace Architecture4

NVIDIA Ampere Architecture NVIDIA Turing™ NVIDIA Volta™ NVIDIA Pascal NVIDIA GPU with Compute Capability SM 5.0 - SM 8.7

CUDA Toolkit 11.8 or later
cuDNN 8.3.x NVIDIA GPU with

Compute Capability 3.5 or higher

CUDA Toolkit 11.5 or later CUDA Driver 450.80.02 or later x86_64 Ubuntu 20.04 LTS 5.4.0-58-generic 7.5.0 2.31
Ubuntu 18.04 LTS 4.15.0-101-generic 7.5.0 2.27
CentOS 7 3.10.0-1160.45.1 8.3.0 2.17
cuDNN 8.2.x CUDA Toolkit 11.2 or later CUDA Driver 450.36.06 or later Ubuntu 20.04 LTS 5.4.0-58-generic 7.5.0 2.31
Ubuntu 18.04 LTS 4.15.0-101-generic 7.5.0 2.27
CentOS 7 3.10.0-1160.45.1 8.3.0 2.17
cuDNN 8.1.x CUDA Toolkit 11.0 or later CUDA Driver 450.36.06 or later Ubuntu 20.04 LTS 5.4.0-58-generic 7.5.0 2.31
Ubuntu 18.04 LTS 4.15.0-101-generic 7.5.0 2.27
CentOS 7 3.10.0-1160.45.1 8.3.0 2.17
cuDNN 8.0.x CUDA Toolkit 11.0 or later CUDA Driver 450.36.06 or later Ubuntu 20.04 LTS 5.4.0-58-generic 7.5.0 2.31
Ubuntu 18.04 LTS 4.15.0-101-generic 7.5.0 2.27
CentOS 7 3.10.0-1160.45.1 8.3.0 2.17
cuDNN 7.6.x NVIDIA GPU with

Compute Capability 3.0 or higher

CUDA Toolkit 10.0 or later CUDA Driver 410.48 or later Ubuntu 18.04 LTS 4.15.0-101-generic 7.3.0 2.27
CentOS 7 3.10.0-1160.45.1 8.3.0 2.17
cuDNN 7.5.x CUDA Toolkit 9.0 or later CUDA Driver 384.81 or later Ubuntu 16.04 LTS 4.15.0-101-generic 5.4.0 2.23
CentOS 7 3.10.0-1160.45.
Retrieved from "http://wiki.hpcmate.com/index.php?title=CuDNN&oldid=256"