GTC 2020 S21455
Presenters: Minseok Lee,NVIDIA ;Zehuan Wang ,NVIDIA
Abstract
We’ll introduce HugeCTR, an open source, GPU-accelerated click-through rate training framework. CTR estimation is a large-scale learning problem that’s essential to the online advertising and content feed industry. Unlike traditional deep learning tasks, CTR training requires terabyte-sized embedding tables, massive I/O, and dynamic feature vocabulary. HugeCTR addresses all of these challenges while achieving 12-44x speedup over existing CPU and hybrid CPU/GPU solutions. HugeCTR is an end-to-end training solution with all computations performed on the GPU, leaving I/O to the CPU. The GPU hash tables support dynamic feature vocabulary. It leverages message-passing interface for multi-node training to support arbitrary large embedding sizes. It also supports mixed precision training, exploiting tensor cores on Volta and its successors.
Watch this session
Join in the conversation below.