Support for GA144

Energy Tutorial

The Terms "Energy" and "Power"

It is important to be clear on the distinction between power and energy, especially when evaluating the implications of GreenArrays chips in systems.

Energy

Energy is the resource that makes it possible to do a given amount of work. It is like money in the bank, a charged battery, a gallon of gasoline, or an equipped and supplied army. Once that amount of work has been done, the energy has been consumed and no more work may be done until more energy has been obtained.

The key concept is that it is the amount of energy that must be consumed in order to do a given amount of work that defines the cost of doing that work. Any system that is not plugged into a wall outlet can only sustain a limited amount of this cost before it has depleted its resources; even a solar-powered device must store energy when the Sun is not available, in order to survive (and do useful work) until the next time the Sun is available.

Power

Power is the rate at which work is done, and thus energy consumed. All other things being equal, to do a given amount of work will consume the same amount of energy if it is done in a microsecond or in an hour; however, the power being applied during that microsecond will have to be 3.6 billion times the power that would need to be applied to do it in an hour.

All things are not completely equal, and there are often inefficiencies in the use of energy at both the extremely fast and extremely slow ends of the scale. These inefficiencies result in the consumption of more energy than would be required to do the work in a more optimal amount of time. At the very fast end, excessive heat can be dissipated and, unless part of the intended work is to behave as a heating element, this heat is wasted and can create the need for still more waste of energy in removing that heat. At the very slow end, effects such as nonlinear friction or semiconductor leakage can lead to inefficiencies.

The key concept is that power is merely a measure of the rate of consumption; it is not something that can be consumed, it is not something that one has a supply of, and thus it is something that must be managed for side effects but not treated as a resource, because it is not a resource.

Units

Power may be measured in watts. On the scale of events inside these chips, nanowatts are often the more practical unit (see nanowatt multiples).

Energy may be measured in watt-hours; one watt-hour is the amount of energy required to do work at an average rate of one watt over one hour.

On the scale of small electronic devices, joules are more useful; one joule is one watt-second. For these chips, a joule is very large. A single add can be on the order of 10 picojoules, so one joule supports roughly 100 billion adds.

GreenArrays Power

Power is minimized by designing circuitry that, wherever possible, moves small charges on and off small capacitive loads through low-resistance transistors. The same principle is carried into I/O so board-level design can preserve those gains.

The resistances and capacitances involved are largely dictated by semiconductor technology. The natural speed of such a system is defined by time constants derived from that technology.

Speed Does Not Necessarily Cost Energy

There are three general ways to slow such a system, and all impose penalties. Gates may be slowed by increasing resistance or capacitance, which increases switching energy. The system may be slowed by clocking, but clock distribution and clocking overhead are high in both device count and energy, and they increase response latency.

Rather than slowing machines below natural speed, the asynchronous model runs when work exists and stops while waiting for data or commands. The computer draws significant power only while executing; while waiting it mostly draws leakage. The energy to stop or start is negligible, on the order of femtojoules.

The result is running fast when needed and stopping immediately when work is done, consuming energy for useful work and near none while idle. In this model, speed itself does not impose extra energy cost; slowing can even consume more.

GreenArrays Energy

The goal is high efficiency: minimizing energy expended per unit of useful work.

Simplicity

Efficiency is achieved by insisting on simplicity in strategy, silicon, software, and system design. A core has fewer than 20,000 transistors, many used for memory and stacks. Internal design changes seek to reduce transistor count and capacitance because every switching element costs energy.

More complex computers generally require more gates to switch per instruction. Published comparisons often cite far lower per-operation energy for these simple cores versus larger architectures. Those savings can still be lost through inefficient software choices or poor system integration.

Design discipline matters at the software and system levels: avoid unnecessary high-cost operations, and avoid board-level choices that add parasitic loss and negate chip-level efficiency.

High Impedance and Low Current

Designs avoid resistors unless necessary, especially where sustained current would turn them into heating elements. Sometimes that cost is unavoidable, but indiscriminate termination of all signals is intentionally avoided.

As a result, most designs avoid termination resistors. Optional weak pull-downs may be enabled around 50k ohms, but high-speed signaling generally avoids them. Input pads present very light loading, with picofarad-scale capacitance and very high impedance to rails.

Software Intensive I/O vs Glue Chips

(... to be continued)