Architecture Support for FPGA Multi-tenancy in the Cloud

摘要: Cloud deployments now increasingly provision FPGA accelerators as part of virtual instances. While FPGAs are still essentially single-tenant, the growing demand for hardware acceleration will inevitably lead to need methods and architectures supporting multi-tenancy. In this paper, we propose an architecture space-sharing devices among multiple tenants in cloud. The proposed implements a network-on-chip (NoC) designed fast data movement low footprint. Prototyping on Xilinx Virtex Ultrascale + demonstrated near specification maximum frequency on-chip high throughput instance access accelerators. We demonstrate similar performance compared single-tenant deployment while increasing utilization (we achieved $6 \times$ higher with our case study), which is one major goals virtualization. Overall, NoC interconnect about $2 than state-of-the-art bandwidth 25.6 Gbps.