As Moore's law slows down, CPUs offer less annual incremental performance per watt, while demand continues to increase. To maintain economic computational performance in the face of increasing demand, data centers are deploying hardware accelerators specialized for particular tasks, e.g., machine vision, video compression, etc. Creating custom application specific integrated circuits (ASICs) for specialized tasks is expensive and requires a highly skilled team.

The techniques of this disclosure describe a workload-identifying process that automatically identifies computational hotspots amenable to hardware acceleration. FPGA designs for such workloads may be machine generated. The FPGA design is tested, and if found worthy by the measure of economic return-on-investment (RoI), taped out as an ASIC. The RoI is fed back to the workload-identifying process, which uses such feedback to improve identification of economically relevant computational hotspots.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.