[doc] performance: Efficient Software Prebuilds (#15147)

* Efficient Software Prebuilds

* improve

docs/source/performance.mdx

@@ -606,6 +606,18 @@ And for Pytorch DeepSpeed has built one as well: [Mixture of Experts](https://ww
 
 
 
+### Efficient Software Prebuilds
+
+PyTorch's [pip and conda builds](https://pytorch.org/get-started/locally/#start-locally) come prebuit with the cuda toolkit which is enough to run PyTorch, but it is insufficient if you need to build cuda extensions.
+
+At times it may take an additional effort to pre-build some components, e.g., if you're using libraries like `apex` that don't come pre-compiled. In other situations figuring out how to install the right cuda toolkit system-wide can be complicated. To address these users' needs PyTorch and NVIDIA release a new version of NGC docker container which already comes with everything prebuilt and you just need to install your programs on it and it will run out of the box.
+
+This approach is also useful if you want to tweak the pytorch source and/or make a new customized build.
+
+To find the docker image version you want start [here](https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/), choose one of the latest monthly releases. Go into the release's notes for the desired release, check that the environment's components are matching your needs (including NVIDIA Driver requirements!) and then at the very top of that document go to the corresponding NGC page. If for some reason you get lost, here is [the index of all PyTorch NGC images](https://ngc.nvidia.com/catalog/containers/nvidia:pytorch).
+
+Next follow the instructions to download and deploy the docker image.
+
 
 ## Contribute