Per Torstein =?iso-8859-1?Q?R=F8ine?= ([email protected])
15 Jul 1999 19:07:03 +0200

"Johnson, David" <[email protected]> writes:
> It's interesting to experiment with just how long a real flip-flop in a real
> lab can actually be "persuaded" to stay metastable when deliberately nudged
> into such a state in a highly controlled manner. I vaguely recall that
> Intel may have done such experiments a long time ago and succeeded, in an
> extremely low-noise environment with extremely finely adjustable timing, in
> getting a simple "74S" series flip-flop to stay metastable for four hours
> (!) before it either snapped back to a valid state or the experiment was
> terminated.

I find this extremely hard to believe. In order to be useful with
asynchronous inputs, flip-flops should have internal positive feedback
(dynamic ones usually don't). Then the probability of metastability
decreases exponentially with time, measured in time constant units. If
I recall correctly, this time constant is the reciprocal of the
flip-flop's gain bandwidth product, or in other words the feedback
loop delay divided by the open-loop gain. The time constant is
probably a few picoseconds in state-of-the-art CMOS.

When you connect two flip-flops after each other, and only use the
last output, you essentially double the available time for positive
feedback. As the relationship is exponential, this could mean an
improvement from a MTBF of once per second to once in the lifetime of
the universe. I would expect that Hamlet manuscript from those famous
apes before seeing four _seconds_ long metastability in a usable

Best regards,

Per Torstein Roeine                            email: [email protected]
University of Oslo                             phone: +47 22 85 24 52
Dept. of Informatics, Microelectronics Group     fax: +47 22 85 24 01
Box 1080 Blindern, N-0316 OSLO, NORWAY

**** To unsubscribe from si-list: send e-mail to [email protected] In the BODY of message put: UNSUBSCRIBE si-list, for more help, put HELP. si-list archives are accessible at ****