Related Session in GTC 2022
Merlin HugeCTR: 由GPU加速的推荐系统训练和推理 [S41352] - Minseok Lee | NVIDIA MERLIN 推荐系统团队高级经理
Merlin HugeCTR: 使用GPU Embedding 缓存的分布式分层推理参数服务器 [S41126] - Yingcan Wei, Fan Yu, Matthias Langer | NVIDIA
Building and Deploying Recommender Systems Quickly and Easily with NVIDIA Merlin [S41119]
– Even Oldridge, Senior Manager, Merlin Recommender Systems Team, NVIDIA
Getting started
HugeCTR @ NVIDIA: developer.nvidia/nvidia-merlin/hugectr
HugeCTR @ GitHub: github/NVIDIA-Merlin/HugeCTR
Success stories
HugeCTR o Leading Design and Development of the Advertising Recommender System at Tencent:
Resources An Interview with Xiangting Kong
o Meituan /
Optimizing Meituan’s Machine Learning Platform: An Interview with Jun Huang
Learn more about HugeCTR
o Accelerating Embedding with the HugeCTR TensorFlow Embedding Plugin
o HugeCTR Series Part 1:
Scaling and Accelerating large Deep Learning Recommender Systems
(CN) HugeCTR 系列第 1 部分：扩展和加速大型深度学习推荐系统
o HugeCTR Series Part 2:
Training large Deep Learning Recommender Models with Merlin HugeCTR’s Python APIs
(CN) HugeCTR 系列第 2 部分：使用 Merlin HugeCTR 的 Python API 训练大型深度学习推荐模型
o HugeCTR Parameter Server Series Part 1: Introduction to Hierarchical Parameter Server
We are Hiring(Full Time & Intern) : C++ Engineer, CUDA Engineer, Recommendation System Algorithm Researcher
Please email your Resume to: sh-recruitment@nvidiaMERLIN HUGECTR: GPU-ACCELERATED RECOMMENDER
SYSTEM TRAINING AND INFERENCE
JERRY SHI RECOMMENDERS
THE PERSONALIZATION ENGINE OF THE INTERNET
DIGITAL CONTENTE-COMMERCE SOCIAL MEDIA DIGITAL ADVERTISING
4.3B Watch Videos Online3.7B Shop Online 4.3B Active Users 4.7B Internet Users
“Already, 35 percent of what consumers purchase on Amazon and 75
percent of what they watch on Netflix come from product
recommendations based on such algorithms.”
Source: McKinsey
3NVIDIA Merlin addresses
Recommender System challenges
NVTabular HugeCTR Triton
ETLData Loading Training Inference
Embedding tables of High throughput to
Using common item- large deep learning rank more items is
Pipelines are slow
by-item loading can recommenderdifficult while
Challenge and complex
be slowsystems can exceed maintaining low
memorylatency
Asynchronous and Easy data and
GPU-accelerated GPU-accelerated High throughput,
and easy-to-use ETLmodel parallel low-latency
Solution dataloader for training allow to
pipelines prepares PyTorch and production
datasets in minutesscale TB sizedeployment
TensorFlow/Keras embeddings
NVIDIA Merlin is an open-source library to deploy recommender systems end-2-end
4AGENDA
HugeCTR Overview
HugeCTR Inference
HugeCTR Sparse Operation Kit
5