Skip to content

Accessing TensorCores in Keras / CuDNN autotuner #9825

@ASvyatkovskiy

Description

@ASvyatkovskiy

In TensorFlow, I can specify to use the CUDNN autotuner by setting: os.environ['TF_CUDNN_USE_AUTOTUNE'] = "1"
(currently enabled by default)”, which improves performance significantly over: os.environ['TF_CUDNN_USE_AUTOTUNE'] = "0"
(I am currently running on TensorCores on Volta GPUs).

However, I am unable to access this performance improvement when running pure Keras. Setting os.environ['TF_CUDNN_USE_AUTOTUNE'] = "1" does not have any effect. Is the autotuner disabled by default in Keras? How can I enable it?

Similar issue: #9321

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions