In principle, digital adiabatic circuits can achieve an internal energy dissipation per operation that can be made as low as is needed, limited only by the leakage characteristics of the device technology used. For MOSFET technologies, the relevant quantity is the maximum transistor Ion/Ioff ratio. Fortunately, the semiconductor industry expects to maintain an Ion/Ioff in the millions for at least the next ten years [ITRS], which leads to a comparable factor of power/performance reductions that will be achievable internally in adiabatic circuits.
Unfortunately, the reduction in total system power dissipation achievable in previous adiabatic computing systems was severely limited by the achievable energy efficiency (characterizable as a Q parameter) of the available resonant power supply designs that were used to generate the AC power signal that is required in order to drive adiabatic logic circuits. [] Typical supply designs suffered from energy losses in the power switches that were needed in order to produce the right shape of logic signal. [] The inefficiencies inherent in these power supply designs (Q's typically in the 10s or lower) meant that adiabatic systems based on them could only achieve around a factor of 10 or less in total energy savings.
There is therefore an evident need for radically different and more efficient resonant power supply designs, if adiabatic technology is to fulfill its potential as a low-power technology.
Recently, we hit on a new design concept for an adiabatic power supply. The radical aspect of this design is that it is based on a mechanical (MEMS) oscillator, as opposed to the purely electronic oscillators that were the basis of earlier approaches. The advantage of this approach is that MEMS structures built without sliding contacts can offer Q's that are in the tens of thousands, thereby enabling comparable factors of energy savings if they are used to drive adiabatic circuits. The right signal waveform shape is attained by harnessing the geometry of the oscillating structure, so that no dissipative switching action is needed in the design.
In ordinary high-performance (as opposed to low-power) digitial systems, MEMS oscillators are inappropriate as clocking mechanisms, since they cannot achieve the high frequencies required. But in adiabatic circuits, our requirements are different: we have to run at relatively low frequencies anyway, in order to achieve a reduction in energy dissipation per operation.
For example, in one typical scenario, an adiabatic circuit would be running at a frequency that is a factor of 1000 slower than a comparable high-performance circuit, in order to achieve an order-1000 factor in energy savings. If the high-performance circuit would have run at 1 GHz, then the adiabatic circuit only needs to run at 1 MHz. Oscillation frequencies on the order of hundreds of kilohertz or megahertz can be readily achieved in MEMS structures today. Even lower-frequency designs could yield an even greater energy savings, for applications that need it.
We will start with our presently-sketched design concept. Dr. Frank and one of Dr. Nishida's graduate students will visit Sandia labs and take the training course that they offer, to learn how to use the design tools for Sandia's commercially available MEMS process. They will then return to UF, and the student will work on refining the MEMS design and characterizing its relevant parameters through analysis and simulation, with consultation with Dr. Ngo concerning power systems design issues. When the design is complete, Sandia will fabricate it in their facility.
Meanwhile, a second student, with guidance from Dr. Frank and resources borrowed from other colleagues in the ECE department (such as Dr. Bill Eisenstadt, in VLSI) will design a simple VLSI adiabatic test circuit (e.g. shift register) intended to exercise the power supply. When completed, this design will be fabricated through MOSIS.
In the meantime, a third student (possibly an undergraduate) would help construct an experimental testing setup, which will include the power-supply to test-chip interface, electrical equipment to measure the total power supplied to the system, IR imaging to pinpoint where dissipation is occurring, and sensitive thermoelectrics-based calorimetry equipment (based on techniques described in []) to precisely measure extremely small levels of dissipation (down to the nanowatt level) within subsystems.
When all components are completed, all parties will help to integrate them into the test setup, after which an extensive set of experiments will be run to accurately characterize the levels of dissipation in the power supply and logic subsystems, and determine the reasons for any discrepancies with theoretical predictions. The MEMS part would be designed to be frequency-tunable, or at least to operate at any of several predetermined frequencies, so that the functional relationship between frequency and dissipation can also be characterized and compared with theoretical expectations.
The results of the research will of course be presented at relevant conferences, and recorded in journal articles. If the project is successful, then it will open the door to a number of possible follow-on research projects (involving a lot of circuits and systems architecture work) to create practical adiabatic systems using the MEMS power supply technolgy, for a variety of low-power computing applications.
This project is very suitable for the NSF's SGER (small grants for exploratory research) program, because it represents: