Accelerating Real-Time Video Analytics on a Heterogeneous CPU and FPGA Platform
Databricks via YouTube
Overview
Syllabus
Why Real-Time ?
Real-Time Insights
Existing Real-Time Analytics Solution
Real-Time Video Analytics
Main Phases of Video Analytics Pipeline
Architecture of CPU based Pipeline
Inference Phase
From DStreams to Structured Streaming
Custom Connector: reading from custom data source
Code Snippet
Deep Learning Inference
Performance of CPU based solution
Challenges with existing software-based solutions
Hardware Accelerators - Alternate Solution
Heterogenous CPU+FPGA based pipeline
Performance of CPU+FPGA based solution
Distributed System Configuration
Demo Application
Summary
Megh Solution Stack: reduces complexity of programming FPGA
Taught by
Databricks