Large-Scale Modular Acceleration for AI, Search, and In-Memory Computing

We propose a new modular FPGA-based server for specialized cloud computing in AI and search with extreme scalability, flexibility, and integrated performance. One of the key features in the server is a low-level software abstraction layer that partitions a large modular array of FPGAs into groups allocated for accelerating various workloads such as computer vision inference, elastic search, and in-memory computing. The FPGAs are configured dynamically with high-abstraction bit-streams for complete offload acceleration. The software layer provides low-latency conduit between and within the FPGA groups, creating real-time performance for connected application such as taking a captured image (from a camera) to search through documents and retrieve geoinformatics data from an in-memory database for recommendations and security.