2
\$\begingroup\$

I'm looking for references to a 1990s research project in Japan involving Xilinx FPGAs.

When a Xilinx FPGA from the late '90s was run below specified power, its behavior would change in unpredictable ways.

(That's because of the difficulty of forcing each signal and each state to be either HIGH or LOW, as opposed to the ease with which they assume values neither high nor low.)

As of 1997-1998, a (British) researcher working in Japan on the Fifth Generation Project was USING this effect as the focus of his research. He found that supposedly-intractable computing problems can be solved by evolving FPGA to exploit these chip-level anomalies. (Specifically, a device that supposedly can hold only eight bits could be "trained" to distinguish between a 1Kb and a 3 Kb input signal. The speculation was that irregularities in manufacturing were being "exploited" by the process of evolving the instrument, to allow it to 'count' beyond 256... a bit like steganography, perhaps, but as an emergent and not pre-programmed effect.)

Does anybody remember that researcher's name? I've searched journal archives, and found no leads.

My colleagues in Complexity and Cybernetics research are anxious to get details on this! (You could ponder the phrase "ultra-cheap quantum computing," to understand our excitement.)

Thanks for any leads at all!!

\$\endgroup\$
5
  • \$\begingroup\$ I recall reading a story in Discover Magazine in the late 90s about using genetic algorithms to program FPGAs that matches some of the details of your question. Was it perhaps Adrian Thompson or Inman Harvey? \$\endgroup\$
    – hobbs
    Sep 22, 2021 at 3:47
  • \$\begingroup\$ Sounds like this? damninteresting.com/on-the-origin-of-circuits Configuration files that will only run on the FPGA they were developed on and no others with strange dead end or isolated islands logic where if removed would prevent the design from working properly. \$\endgroup\$
    – DKNguyen
    Sep 22, 2021 at 3:54
  • \$\begingroup\$ You might find the research into "Physically unclonable functions" relevant, much the same stuff, exploiting timing differences, and logic elements as analogue to produce something unique to a specific part. I do remember reading about the work you mention. but cannot remember a real reference. \$\endgroup\$
    – Dan Mills
    Sep 22, 2021 at 14:10
  • \$\begingroup\$ I believe this may have been specific to the long-dead Xilinx XC6200 FPGA family. \$\endgroup\$
    – user16324
    Sep 22, 2021 at 15:33
  • \$\begingroup\$ Thanks, Everybody! This is exactly what I've been trying to track down for some years. And this gives me an opening to some folks I've been hoping to engage with - Phil Husbands, Hugo de Garis - for some tangential projects. This all goes to show: When you have a question, ask somebody with KNOWLEDGE! \$\endgroup\$ Sep 23, 2021 at 6:40

1 Answer 1

3
\$\begingroup\$

I don't recall anything about undervolting, or anything about Japan, but Adrian Thompson was a British researcher did some work around that time on using genetic algorithms to program FPGAs to solve problems, and one of them was discriminating between tones at two different frequencies.

The solutions that the algorithm came up with were "weird" and relied on undefined behaviors that couldn't be replicated in simulation, and frequently failed to work when loaded onto another "identical" chip, or when the ambient temperature changed. It was doing things like creating unclocked free-running oscillators out of inverter loops, and inductively coupling signals between paths that were (as far as the digital logic was concerned) not connected at all.

Evolving A Conscious Machine, Discover Magazine, June 1998.

On the Origin of Circuits, Damn Interesting, June 2007 (spotted by @DKNguyen).

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.