Quality-Time Tradeoffs in Component-Specific Mapping: How to Train Your Dynamically Reconfigurable Array of Gates with Outrageous Network-delays

How should we perform component-specific adaptation for FPGAs? Prior work has demonstrated that the negative effects of variation can be largely mitigated using complete knowledge of device characteristics and full per-FPGA CAD flow. However, the cost of per-FPGA characterization and mapping could be prohibitively expensive. We explore light-weight options for per-FPGA mapping that avoid the need for a priori device characterization and perform less expensive per FPGA customization work. We characterize the tradeoff between Quality-of-Results (energy, delay) and per-device mapping costs for 7 design points ranging from complete mapping based on knowledge to no per-device mapping. We show that it is possible to get 48-77% of the component-specific mapping delay benefit or 57% of the energy benefit with a mapping that takes less than 20 seconds per FPGA. An incremental solution can start execution after a 21 ms bitstream load and converge to 77% delay benefit after 18 seconds of runtime.