Frequency scaling means adjusting the clock frequency to a subsystem.

Frequency scaling is software controlled by updating divider ratios. Adjustment can be instant, but reconfiguring PLL has inertia (say 1 millisecond). Let's adjust the clock frequency (while keeping VCC constant for now). Does this help ?

Compare dynamic frequency adjustment with with clock and supply gating:

Clock Gating | Supply Gating | Frequency Adjustment | |

Control: | automatic | various | software |

Granularity: | register / FSM | larger blocks | macroscopic. |

Clock Tree: | mostly free runs | turned off | slows down. |

Response time: | instant | 2 to 3 cycles | instant (or ms if PLL adjusted) |

Proportionally vary voltage: | not possible | n/a | yes. |

Two potential strategies (Zeno: Tortoise v Achilles):

- Compute quickly and then halt.
- Compute more slowly and finish just in time.

To compute quickly and halt we need a higher frequency clock but consume the same number of active cycles.

So the work-rate product, af, unchanged, so no power difference ?

Actually un-stopped regions (clock trees etc.) consume power proportional to f.

Until recently: Tortoise is best: keep going steadily and end just in time. (He appeals even more if we can vary the voltage to be just enough to support this clock frequency.)

But, dynamic clock gating still also useful for: bursty, localised activity (which is the general case).

20: (C) 2008-13, DJ Greaves, University of Cambridge, Computer Laboratory. |