# Channel last enablement in pytorch for convolution

**URL:** <https://dev-discuss.pytorch.org/t/channel-last-enablement-in-pytorch-for-convolution/1880>\
**Category:** Uncategorized\
**Created:** [February 15, 2024, 6:34am UTC](https://dev-discuss.pytorch.org/t/channel-last-enablement-in-pytorch-for-convolution/1880 "2024-02-15T06:34:33Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![Aruna](https://avatars.discourse-cdn.com/v4/letter/a/71e660/32.png) [@Aruna](https://dev-discuss.pytorch.org/u/Aruna)\
**Post date:** [February 15, 2024, 6:34am UTC](https://dev-discuss.pytorch.org/t/channel-last-enablement-in-pytorch-for-convolution/1880/1 "2024-02-15T06:34:33Z")

</div>

With reference to [BKM for PyTorch CPU Performance (github.com)](https://gist.github.com/mingfeima/363a9ab850be54d5837f9cc542ad2b38), the vision based models are faster in NHWC than NCHW. Instead of having API level enablement, created PR [Enable Optimized Performance Boost for 2D Forward Convolution with Channel Last by akote123 · Pull Request #117845 · pytorch/pytorch (github.com)](https://github.com/pytorch/pytorch/pull/117845) to make NHWC default for convolution.  
Please share view on defaulting NHWC for Convolution Op.