Dan Stuart, Penguin Computing
gtc-dc 2019
We’ll show how to join traditional computer vision and AI inference across multiple models using a streaming, asynchronous, high-throughput workflow. The Hybrid Task Graph Scheduler and TensorRT perform scalable, single node, multi-GPU object detection, classification, and regression. This approach targets automated whole slide microscopy operating under tight time constraints with a size of 100,000 by 50,000 pixels. The end-to-end workflow starts with a microscope scan and finishes with a populated database that contains quality assurance metrics, object detection, and more. This generalizable workflow applies three independent TensorRT models with dependency management concurrently across all available GPUs within a single computer.