Low-Energy Computing for
Implantable Medical Devices

MEDG Seminar
February 21, 1996

As experts in the physics of computation have known for some time, there is no lower bound on the amount of energy dissipation needed to perform computation; in principle, the energy can be made as small as desired. Unfortunately, in traditional computers, CPUs dissipate a substantial amount of energy per computational operation, limiting the effectiveness of computers in severely energy-limited (e.g., battery-powered) contexts.

However, recent developments in digital circuit design have shown that it is possible to create so-called ``adiabatic'' digital circuits in which the energy per operation can actually be made as low as desired. The main drawback is that individual operations must be performed more slowly. Our group is creating a general-purpose CPU chip based on this technology. What would the applications of such a device be?

One context in which low energy consumption might be particularly desirable is the case of a battery-powered device to be implanted long-term in the human body, since we might imagine that a need for frequent recharging or replacement would be very undesirable to the patient. With the new technology, an implant with a tiny battery could perform essentially as much computation as desired without needing a fresh input of energy. But what sorts of implants might need to do large amounts of computation?

In this talk, I will briefly outline the technological tradeoffs afforded by the new technology, but then I would like to spend most of the time in an open discussion with the audience about what some possible uses of the technology in medical implants might be. I would like audience members to try to think up some possible applications before coming to the talk. To get you started, here are some initial ideas, some of which are pretty far-out. I hope you can help me come up with some more ideas that are closer to reality.

1. Monitoring device to frequently measure e.g. blood-sugar levels and compute & store long-term statistics & patterns, for later communication (via radio or removal).
2. Device as above, but also with a decision-making and planning component to decide, e.g., when to release a stored drug.
3. Artificial retina that performs needed preprocessing of input data before sending to optic nerve. (Run on light energy, without battery?)
4. Other neural-interface implants to replace the computational role of damaged ganglia, etc.

Keep in mind that although we can perform computations with arbitarily little energy, this says nothing about the energy requirements for sensor/effector operations. Therefore, applications that require lots of computation relative to the amount of interaction with their environment are most convincing as roles for the new technology.

I look forward to an interesting discussion.

Accompanying Paper, 12 pages, 120K, Postscript.

Transparencies, 10 pages, 12K, Postscript.