Title: Composable FPGA Overlays for Hardware Acceleration in the Cloud

Abstract: Hardware accelerators have become an important asset in scaling computing as traditional Von Neumann architectures have failed to deliver gains from transistor scaling. GPUs, capable of highly optimised matrix and tensor operations, have driven the growth in machine learning in the last decade. FPGAs offer additional flexibility for irregular applications. However, these accelerators are typically slaves to host servers, which entails an additional offload overhead. Exploiting the interconnect flexibility of FPGAs allows them to be used independently of host servers, enabling low-overhead composition for distributed acceleration. We review the benefits of FPGAs, current approaches to deployment, and our proposal for serverless FPGA accelerators.