Dive deep into the new TensorFlow Embedding Plugin which is introduced by HugeCTR. In HugeCTR, we implemented several efficient embedding layers, which can leverage multiple GPUs to do the training. Those embedding layers support model parallelism and keep trainable parameters on GPU memory so that the training process is fully accelerated by GPU. Compared with TensorFlow embedding layer, it can achieve 10 times better performance on 8 A100 GPUs. To make it easier to use, we integrate those efficient embedding layers as TensorFlow Plugin so that you can utilize them in your existed TensorFlow code snippet. In this session, we will introduce how those embedding layers works, how to utilize those embedding layers with TensorFlow and its training performance.