Scalable OpenCL FPGA Computing Evolution

Abstract

Multi-FPGA acceleration has already shown several orders of magnitude speedups in today's analytics, scientific, financial, image processing and machine learning applications. However, the scalability of FPGA-based HPC demanded by the future avalanche of data and network traffic is still in its infancy. OpenCL memory model holds the promise of straightforward and pragmatic scalability of FPGA computing beyond single FPGA within the framework of today's OpenCL FPGA compilers. Scientific Concepts International develops novel Smart Cell Interconnect (SCI) optimized for OpenCL global memory, streaming data accesses and network packets encapsulated into switched cells. Recent advances in Open Source tools based on polyhedral model and IR enable development of source-to-source coarse grain code and data partitioning of the HPC workloads written initially in OpenCL and followed by C/C++, SYCL. Evolution of the multi-FPGA partitioning tools and SCI interconnect IP will enable true scalability of the computing fabric of arrays and clusters of FPGAs. Cloud FPGA computing, fog computing at the source of the generated data as well as fusion of networking, security, and computing are addressed by our architecture, partitioner tools, and future product roadmap.