GTC 2020: Tensor Core Performance on NVIDIA GPUs: The Ultimate Guide
After clicking “Watch Now” you will be prompted to login or join.
Click “Watch Now” to login or join the NVIDIA Developer Program.
Tensor Core Performance on NVIDIA GPUs: The Ultimate Guide
Valerie Sarge, NVIDIA
Learn what's needed to achieve optimal performance on NVIDIA Tensor Core GPUs, including the brand-new A100 GPU based on the NVIDIA Ampere architecture. We'll review the fundamentals of GPU performance, explain how Tensor Core-accelerated operations work, and use this knowledge to infer how to structure and size neural network operations (layers) to achieve ideal performance. We'll also provide a cheat sheet of Tensor Core performance guidelines. The talk aims to provide tools to understand why neural networks perform a certain way on Tensor Core GPUs and to enable changes to network architecture to further improve performance.