Managing Static (Leakage) Power - Computer Architecture Techniques for Power-Efficiency - page 167

cache line is dead. Second, the problem of lost opportunity to save leakage waiting for the decay
interval to pass (in high-leakage mode) is solved by using the drowsy mode while waiting. The
hybrid approach works well assuming a constant temperature. Its trade-offs, however, change
as temperature affects leakage.
5.3.4 Temperature
Although the dependence of leakage to temperature is not accounted for in the work of Meng
et al., clearly the trade-off between drowsy and decay modes in a hybrid scheme cannot be
static. The higher the temperature the more valuable the decay becomes, because it saves more.
In contrast, at low temperatures, what matters most is to minimize dynamic energy penalty and
performance loss; this argues in favor of the drowsy mode.
This temperature-driven trade-off is explored in a temperature-adaptive scheme [ 129 ].
The leakage control mechanism is a hybrid drowsy + decay scheme. The decision on how long
to wait to enter the decay mode (i.e., how long is the decay interval) depends on the relative
strength of the leakage power to dynamic power. 16 The dynamic power component is affected
by the number of decay-induced misses which is a function of the decay interval.
At high temperatures, where leakage is relatively strong, the decay mode can be aggres-
sively engaged since, even with an increase in dynamic power overhead, it maximizes the overall
power savings. But when leakage currents are weak, the dynamic power overhead can dominate,
regardless of the amount of saved leakage.
Temperature-driven adaptation consists of using a timing mechanism to stretch the decay
interval at lower temperatures. This diminishes the leakage reduction from the decay mode but
also minimizes its dynamic power overhead, leading to an overall reduction in power over a
wide range of temperatures. The drowsy interval is not changed and is fixed at all temperatures.
While there are several ways to control the decay interval at various temperatures (for
example, using a thermal sensor to set the global counter of a hierarchical counter mecha-
nism), adaptation is based on the decay of quasi-static 4T memory cell. The decay interval
is measured as the time it takes leakage to discharge a 4T cell. In other words, the strength
of the leakage itself determines the decay interval. The key in this idea is that the discharge
characteristic of a 4T memory cell, at various temperatures is ideally suited to be used as the
decay interval. The resulting scheme consistently outperforms the decay mode or the drowsy
mode alone and approaches the performance of a hybrid scheme with oracle decay interval
selection [ 129 ].
16 This technique assumes that temperature alone is the factor that changes the relative strength of leakage at runtime
but dynamic scaling of the threshold voltage and the supply voltage could also have major effects.