CellOpt: Power and Delay Optimization

"Finally, we got the flip-flop that you optimized, working on the silicon. Our chip became functional. We gained about 30% on cell area, and that turned out about 4-5% gain on total silicon area utilization. Perhaps we have been too cautious and didn't use them in high frequency blocks." - Taji Isik, ICS

$\includegraphics{cellopt.ps}$

Power management is the most significant design challenge facing IC designers today. There are widely used system level techniques like low power modes, clock gating techniques and custom low power circuits to address the problem. However, the situation is desperate enough to take extreme measures like special processes and very low voltages, power versus performance trade-offs, etc.

CellOpt is a dynamic circuit level power/timing optimizer designed to minimize the power dissipation of standard build blocks like logic gates and sequential elements, with optional timing constraints, by properly sizing the devices used in the circuit. After all, energy is consumed by these components, and how well they are designed and how well they fit to their particular use on the chip is of paramount importance. Traditionally, cell designers follow certain ratio rules to accommodate drive requirements, but the primary design goals are usually the efficiency of the layout and ease of portability. Power characteristics of the circuits usually fall to the wayside as their proper characterization is beyond the capabilities and time limits of most circuit design engineers.

Short Circuit versus Capacitive Power

There are two competing events at the circuit level. One is the charge storage capacity of the circuit which grows with bigger devices and increases as the circuit switches faster. The other is the short circuit power dissipation: if the transition is fast, this component will be small; if it is slow, it could be significant. The optimization problem balances these two competing events against each other.

SPICE is the most reliable tool.

At the circuit level almost all the capacitors are non-linear. Short circuit current depends very strongly on the device characteristics. The measurement of their precise magnitude and duration is beyond the capabilities of non-spice based tools. Usually the amount of energy is very small, in the range of fractions of pico-Joules. Small errors may throw the solution off the target. There are critical path based device sizing tools which are much faster in giving you an answer although it is usually wrong. Their usefulness drops off very fast as you move away from inverter chains. CellOpt is not a critical path optimizer. It looks at all the paths simultaneously: critical, non-critical, switching and non-switching. The effect of each path is weighted statistically, mimicking its actual realization under typical usage models. Critical path based optimizers look at only a small part of the circuit at a time. As they adjust some paths, they can not control their adverse effects on other parts of the circuit.

Inputs to CellOpt

CellOpt needs to have a spice netlist which is functional. It also requires a functional specification for the cell in ACDL format. The cell should be cleared by SpiceTest before optimization is performed. It also needs capacitive loads for each output, and input rise/fall times for each input. One can specify various timing constraints in the form of upper limits on input/output delays and output rise/fall slew times, and for sequential cells setup times for different input pairs. Individual paths can have different constraints.

The tool has a Perl based setup file to define various choices. Using Perl, one can override various setup parameters and constraints for an individual cell. A problem with pre-layout netlist is the estimation of device parasitics. Proper modeling of the device parameters is important for the reliability of the optimization results. The user is expected to provide a Perl function which returns the device parameters given its length and width. It is also also possible to generate these functions automatically if the circuit layout topology is not going to change and if you can generate the layout automatically. Input switching rates are calculated by CellOpt. The user can increase the toggle weight associated with each inputs.

CellOpt is normally executed by MakeLib. MakeLib verifies the functionality of the circuit using SpiceTest, analyzes its power dissipation modes using StateGen and starts the optimization process. It generates a new netlist with modified device sizes.

Timing versus Power

Timing requirements take precedence over power optimization. Any constraints on the circuit will restrict any possible optimization which can be done. If it can't change the size of the output devices, which are the most critical for power dissipation, emphasis will be placed on the other devices, like smaller devices at the inputs. Depending on the initial circuit size, it is possible to end up with a lower power, faster and smaller circuit. Conflicts between the constraints are resolved using their user assigned weights.

Low Power Strategy

Typically, clocks are the most heavily loaded nets, and consume the most of the energy. Clock network optimization by LowSkew can help reduce the clock power usage without introducing skews. Registers driven directly by clocks could be optimized for low power by CellOpt. These cells are the most likely to toggle and consume energy. Any improvement in their power usage affects the energy consumption. Finally for core logic, building a “fixed-delay” library with CellOpt can reduce the power dissipation without affecting the performance. Usually, standard cells are designed for fixed rise/fall times. However, what counts in a gate level design is the total delay through the cell. By distributing the delay throughout the gate, one can design more energy efficient cells while meeting the performance goals.

$\includegraphics{waveforms.ps}$

Delay versus Power

Faster circuits use more energy. Although this is true, it does not exclude good design practices. Figure 1 shows power waveforms for C1908 from ISCAS benchmark circuits after synthesizing it for zero delay using a commercial library. CellOpt was used on this library to generate the fastest possible worst case pin-to-pin delays. Figure 2 shows the power waveforms generated by PowerTeam after synthesizing the same circuit with the optimized library for best timing. This speeds up slow timing paths, and slows down fast timing arcs, with small change in the average signal delay through the cell. The new implementation is slightly faster, 1.725ns versus 1.706ns, and more energy efficient according to the synthesizer, by about 13%. It is noteworthy that peak power varies by as much as 50% between the two implementations, depending on the inputs. Over a much longer interval, energy use difference is 10% based on PowerTeam simulations. For more significant reductions such as 30%, one needs to use the faster cells on the critical paths, and the slower, more energy efficient cells on the non-critical paths. Such paths are the most common.