Accelerating Real-Time Video Analytics on a Heterogeneous CPU and FPGA Platform

Explore a conference talk on accelerating real-time video analytics using a heterogeneous CPU + FPGA platform. Dive into the challenges of implementing video analytics solutions in industrial sectors, focusing on the need for low latency and cost-effective deployment. Learn about a proposed solution leveraging Spark Structured Streaming and deep learning frameworks on a CPU + FPGA hardware platform, offering 3x performance acceleration and 2x decrease in TCO compared to CPU-only implementations. Discover the key components of the video analytics pipeline, including video stream ingestion, H.264 decoding, image transformation, and inferencing using deep neural networks. Gain insights into optimizing Spark Streaming and deep learning pipelines, accelerating video analytics using FPGAs, and comparing performance benchmarks between CPU and CPU + FPGA configurations. Understand the architecture of CPU-based pipelines, the transition from DStreams to Structured Streaming, and the use of custom connectors for data sources. Explore the challenges of existing software-based solutions and how hardware accelerators like FPGAs can address them. Conclude with a demo application and an overview of the Megh Solution Stack for simplified FPGA programming.