1. Low-Power and Temperature-Aware Compilation for Embedded Processors José L. Ayala Politecnica University of Madrid jayala@die.upm.es

2. Outline Motivation Motivation Goals Goals Low-Power Compilation Low-Power Compilation Temperature-Aware Compilation Temperature-Aware Compilation Ongoing and Future Work Ongoing and Future Work

3. Motivation Continuing advances in semiconductor technology  performance gains (clock rate and ILP) Continuing advances in semiconductor technology  performance gains (clock rate and ILP) Significant increase in power consumption (dynamic and static) Significant increase in power consumption (dynamic and static) Technology scales  future power density 200 W/cm 2 : early consideration of thermal effects are needed Technology scales  future power density 200 W/cm 2 : early consideration of thermal effects are needed

4. Goals Extensive efforts on dynamic techniques: static approaches are needed Extensive efforts on dynamic techniques: static approaches are needed Early thermal characterization with a selectable granularity Early thermal characterization with a selectable granularity Thermal and power management incorporated in the compiler code Thermal and power management incorporated in the compiler code

5. Low-Power Compilation Power-aware compilation for embedded single-core processors Power-aware compilation for embedded single-core processors –Modified register assignment: power reduction in the register file –Register file organized in banks: simple decoding logic –Register assignment improves register locality: registers assigned from same bank

6. Low-Power Compilation Power-aware compilation for embedded single-core processors Power-aware compilation for embedded single-core processors –Voltage scaling technique applied to unused banks –Compiler modifications in gcc compiler –Energy savings up to 75% without performance penalty

7. Low-Power Compilation Power-aware compilation for VLIW- based systems Power-aware compilation for VLIW- based systems –Power-aware register assignment targeting VLIW-based systems –Similar approach than the applied to embedded processors –Compiler modifications in Trimaran compiler

8. Low-Power Compilation Power-aware compilation for VLIW- based systems Power-aware compilation for VLIW- based systems –Simulation framework: CRISP (by IMEC) –Energy savings up to 60% without performance penalty –Modified compiler helps on simplifying over-sized register file in compilation time

9. Low-Power Compilation

10. Temperature-Aware Compilation Two different goals: Two different goals: –Decrease total temperature in the chip: effect on propagation delays, signal integrity and power consumption –Avoid temperature gradients: effect on electromigration and chip damage

11. Temperature-Aware Compilation First step: thermal modeling First step: thermal modeling –Thermal modeling based on power consumption measures: bidimensional RC thermal model through layout area estimations (black box model) –Thermal characterization of hardware modules in the chip area

12. Temperature-Aware Compilation

13. Second step: thermal analysis of source-level transformations Second step: thermal analysis of source-level transformations –Analysis of temperature gradients

14. Temperature-Aware Compilation

15. Second step: thermal analysis of source-level transformations Second step: thermal analysis of source-level transformations –Analysis of temperature gradients –Analysis of global temperature

16. Temperature-Aware Compilation

17. Third step: temperature/power-aware compiler Third step: temperature/power-aware compiler

18. Ongoing and Future Work Compiler-driven reconfigurable logic for low-power consumption: improves register assignment Compiler-driven reconfigurable logic for low-power consumption: improves register assignment Implementation over a real platform (CoolFlux, Lisatek, ACE/COSY…) Implementation over a real platform (CoolFlux, Lisatek, ACE/COSY…) Extension of the number of analyzed code transformations with impact in the thermal behavior Extension of the number of analyzed code transformations with impact in the thermal behavior

19. Ongoing and Future Work Integration of low-power optimizations and temperature-aware transformations in the compiler code Integration of low-power optimizations and temperature-aware transformations in the compiler code

20. Thanks! Questions?