Phase I Project Outcomes Report*
This Small Business Innovation Research (SBIR) Phase I project focuses on enabling energy-efficient customized computing for big data applications in datacenters. The most significant barrier for widespread adoption of FPGA-based customized computing is the difficulty in programming FPGAs. The innovation of this project includes the development of an automated compiler for FPGA accelerator creation, a runtime system for managing accelerator resources in datacenters, as well as a high-performance FPGA acceleration library for selected machine learning algorithms. We developed an initial baseline version of an integrated solution that includes all three components. The solution is also integrated with big data frameworks Apache Hadoop and Spark to enable efficient and transparent utilization of FPGA accelerators in datacenters. Initial results show significant improvement in programming productivity with the compiler and substantial performance advantage and energy savings with the machine learning accelerators managed by the runtime system over CPUs and GPUs. For some of the applications, we demonstrated that several computing servers can be replaced with one server augmented with one or more FPGA acceleration cards, resulting in significant cost and energy reduction. The project further demonstrates the viability and advantages of the FPGA-based customized computing in datacenters against multi-core servers and GPUs in terms of performance and energy-efficiency.
Wide deployment of customized computing technology in datacenters as enabled by the solutions developed in this project can lead to substantial energy savings, significant carbon emission reduction, and more importantly, more sustainable growth of computing infrastructures so that they can better scale in the future to meet the rapidly increasing computing demands as our society embraces further digital revolution in the coming decades. We also shared some of the innovations and results from this project with broad community with two keynote speeches at technical conferences and a forthcoming book chapter, which can lead to growing interest and increased adoption of customized computing in datacenters.
*Any opinions, findings, and conclusions or recommendations expressed in this report are those of the PI and do not necessarily reflect the views of the National Science Foundation; NSF has not approved or endorsed its content.
