A Distributed Approach to Silicon Compilation
TimeWednesday, July 13th2:30pm - 3pm PDT
Location3003, Level 3
Special Session (Research)
DescriptionExtreme hardware specialization is the only viable path for extending the current exponential electronics improvement trajectory indefinitely, but this path is currently blocked by the high engineering effort of chip design. Specialization for the long tail of end user applications will require the creation of completely automated end-to-end compilers with many orders of magnitude improvements in run-time. The enormity of these challenges means single machine execution and monolithic single company efforts are unlikely to be sufficient. In this work, we take a distributed systems approach to compilers, with the goal of creating infrastructure that scales to thousands of developers and millions of servers. Technical contributions include a unified hardware compiler schema, a client/server type distributed execution model, and an open source Python framework designed for warehouse scale execution. An order of magnitude compilation speed-ups have been demonstrated on multiple designs and PDKs compared to traditional make based flows.