The demand for real-time stream processing is increasing rapidly with the explosion of data from the Web, sensors, IoT and mobile devices, and other sources. Enterprises want to process this data as it moves to create business value in areas such as:

• Finance (e.g., fraud detection, risk management)
• Sales and marketing (e.g., customer preferences)
• Operations (e.g., predictive asset maintenance, operational efficiencies)
• Network management (e.g., optimization, location intelligence)

Opensource and proprietary software-only solutions cannot meet this demand, driving the need for hardware accelerators to scale performance and offer lower latencies for certain workloads. Field Programmable Gate Arrays (FPGAs), as part of a heterogeneous CPU+FPGA platform, provide the answer, offering reconfigurable accelerators that deliver better performance at lower latency. Unfortunately, while FPGAs are much better suited for these kinds of streaming applications than GPUs or other accelerators, they pose challenges in hardware and software integration, and require specialized programming expertise.

What is needed is an FPGA-based platform that makes it easy for existing teams to create business value. Specifically:

• CIOs and business unit managers need a platform that uses existing hardware and frameworks to support off-the-shelf and custom solutions that deliver compelling value propositions with increased performance, lower cost, and new features
• Data scientists need a platform with abstraction for high-level coding and rapid development of alternate models
• Developers need a rapid-development solution that provides support for Scala, Python, Java, R, and other existing language expertise

Megh is developing such a platform, utilizing standard server infrastructure from OEMs and ODMs for on-premise private data centers and edge computing centers, or hosted platforms provided by leading CSPs. And we are building an ecosystem of academic and IP partners to further support and expand the platform.

Our advanced, real-time analytics platform provides a comprehensive vertical software stack to abstract the complexity of FPGA programming and management, and support transparent application integration. It is based on libraries running on standard high-volume servers equipped with PCIe-attached FPGA cards with Intel or Xilinx FPGAs. The FPGAs process the streaming data and interact with the application running on the host.

Megh breaks down the real-time analytics pipeline into three phases:

1. The ingest stage, in which streaming data in different formats (structured, time series, video, audio, images, text, etc.) is processed to obtain and prepare a payload for further processing
2. The transform stage, where specific data is extracted, transformed, and loaded (ETL) for analysis
3. The infer stage, which today normally involves using machine learning or deep learning networks to discern patterns or recognize images

Megh implements the complete ingest, transform, and infer pipeline in the FPGA as Sira accelerator function units (AFUs) managed by the Arka Runtime. The platform supports both standard frameworks like Spark, TensorFlow, KX (for financial apps) and custom data analytics frameworks, with machine and deep learning libraries and no code changes to existing applications.