Using a lower precision for an application execution can sometimes enable a runtime to meet the dynamic power and deadline constrains. The consortium invites proposals for development of models and supporting runtime/hardware layers for precision based scheduling on the PULP platform. The current version of PULP supports a transprecision FPU (https://arxiv.org/abs/1711.10374) as well as a selection of benchmarks running on PULP (https://github.com/oprecomp/micro-benchmarks).
The focus of work going forward will be to first enable the dynamic switching of precision on the PULP platform supported by transprecision unit in a way that lowering of precision can provide a higher throughput. Another way to highlight the benefits of transprecision scheduling can be to instantiate heterogeneous PULP cores, that is cores offering different precisions. Once the platform supports that, at run-time the scheduler can switch precision based on statically designed analytical models for a single application or port applications to different PULP cores to meet their real-time constraints.
Further details on the PULP ecosystem are available at: https://github.com/pulp-platform/pulp-sdk
For additional information please contact: Umar Minhas