diff --git a/docs/images/precision_options.png b/docs/images/precision_options.png
new file mode 100644
index 0000000000..269560d80f
Binary files /dev/null and b/docs/images/precision_options.png differ
diff --git a/docs/source/index.rst b/docs/source/index.rst
index 1fde0b0ef3..54dc6e6922 100644
--- a/docs/source/index.rst
+++ b/docs/source/index.rst
@@ -58,6 +58,12 @@ Technical documentation is available at `docs.monai.io <https://docs.monai.io>`_
 
   installation
 
+.. toctree::
+  :maxdepth: 1
+  :caption: Precision and Performance
+
+  precision_performance
+
 .. toctree::
   :maxdepth: 1
   :caption: Contributing
diff --git a/docs/source/precision_performance.md b/docs/source/precision_performance.md
new file mode 100644
index 0000000000..f091d5d178
--- /dev/null
+++ b/docs/source/precision_performance.md
@@ -0,0 +1,39 @@
+# Precision and Performance
+
+Modern GPU architectures usually can use reduced precision tensor data or computational operations to save memory and increase throughput. However, in some cases, the reduced precision will cause numerical stability issues, and further cause reproducibility issues. Therefore, please ensure that you are using appropriate precision.
+
+<!-- Maybe adding Automatic Mixed Precision, Float16 or BFloat16 in the future-->
+
+## TensorFloat-32 (TF32)
+
+### Introduction
+
+NVIDIA introduced a new math mode TensorFloat-32 (TF32) for NVIDIA Ampere GPUs and above, see [Accelerating AI Training with NVIDIA TF32 Tensor Cores](https://developer.nvidia.com/blog/accelerating-ai-training-with-tf32-tensor-cores/), [TRAINING NEURAL NETWORKS
+WITH TENSOR CORES](https://nvlabs.github.io/eccv2020-mixed-precision-tutorial/files/dusan_stosic-training-neural-networks-with-tensor-cores.pdf), [CUDA 11](https://developer.nvidia.com/blog/cuda-11-features-revealed/) and [Ampere architecture](https://developer.nvidia.com/blog/nvidia-ampere-architecture-in-depth/).
+
+TF32 adopts 8 exponent bits, 10 bits of mantissa, and one sign bit.
+
+![Precision options used for AI training.](../images/ai_training_tf32_tensor_cores.png)
+
+### Potential Impact
+
+Although NVIDIA has shown that TF32 mode can reach the same accuracy and convergence as float32 for most AI workloads, some users still find some significant effect on their applications, see [PyTorch and TensorFloat32](https://dev-discuss.pytorch.org/t/pytorch-and-tensorfloat32/504). Users who need high-precision matrix operation, such as traditional computer graphics operation and kernel method, may be affected by TF32 precision.
+
+Note that all operations that use `cuda.matmul` may be affected
+by TF32 mode so the impact is very wide.
+
+### Settings
+
+[PyTorch TF32](https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices) default value:
+```python
+torch.backends.cuda.matmul.allow_tf32 = False # in PyTorch 1.12 and later.
+torch.backends.cudnn.allow_tf32 = True
+```
+Please note that there are environment variables that can override the flags above. For example, the environment variables mentioned in [Accelerating AI Training with NVIDIA TF32 Tensor Cores](https://developer.nvidia.com/blog/accelerating-ai-training-with-tf32-tensor-cores/) and `TORCH_ALLOW_TF32_CUBLAS_OVERRIDE` used by PyTorch. Thus, in some cases, the flags may be accidentally changed or overridden.
+
+We recommend that users print out these two flags for confirmation when unsure.
+
+If you are using an NGC PyTorch container, the container includes a layer `ENV TORCH_ALLOW_TF32_CUBLAS_OVERRIDE=1`.
+The default value `torch.backends.cuda.matmul.allow_tf32` will be overridden to `True`.
+
+If you can confirm through experiments that your model has no accuracy or convergence issues in TF32 mode and you have NVIDIA Ampere GPUs or above, you can set the two flags above to `True` to speed up your model.