[1/3] Software Systems on DNN Inference and Training

[widiba03304]

Date: 2022.01.03
Presenter: Minjae Kim
Paper: No specific paper for this presentation.

I'll introduce the holistic ecosystem to deep learning with a top-down approach. Due to the time constraint, I cannot explain every detail of each topic so expect to give a brief introductory talk. I will focus on the key optimization techniques used in the commodity software. All the materials will be written in English but I'll speak in Korean.

Table of contents

• DNN applications (Autonomous driving, VR/AR, Chatbot, …)
• DNN models (CNN, RNN, Transformer, …)
• Model specifications (ONNX, …)
• High-level programming models (PyTorch, TensorFlow, JAX, …)
• Execution engines (TensorFlow Lite, MNN, NVIDIA TensorRT, …)
• Deep learning compilers (XLA, TVM, …)
• DNN Serving systems (TF Serving, NVIDIA Triton, …)
• Vendor libraries (cuDNN, Qualcomm SPNE, …)
• Low-level programming models (CUDA, OpenCL, …)
• Low-level programming languages and compilers (PTX, SPIR-V, …)
• Deep learning hardware (GPU, TPU, NPU, …)