Increased tendency towards multi-product platform chips means large functional blocks on silicon may be off for complete product lifetime.
The `dark silicon' future scenario implies all chips must be mostly powered off.
Battery powered devices will also use macro-scale block power down (e.g. the audio or video input and output subsystems).
Dynamic clock gating can be fairly fine grain, but we can also turn off power to sections of logic (coarser grain).
Use power gating cells in series with supply rails (using gound rail is more area efficient).
And use signal isolation and retention cells (t-latches) on data input and output nets.
No register and RAM data retention in block while off.
Dynamic power gating techniques typically require some sequencing: several clock cycles to power up/down a region and enable/disable isolation gates.
Originally, power off/on controlled by software or top-level input pads. Today a dedicated microsequencer might control 100 regions in a subsystem.
Sometimes power off a whole chip except for a one or two RAMs and register files and perhaps DRAM refresh timer.
Can also retain volatile contents using a lower supply while clock is off (CMOS RAM data holding voltage).
|18: (C) 2008-13, DJ Greaves, University of Cambridge, Computer Laboratory.|