[time-nuts] Metastability

[Magnus Danielson]

Hi Rick,

On 2019-01-21 16:31, Richard (Rick) Karlquist wrote:
> I designed a state machine running on a 32 MHz clock
> to resync 4 MHz clock edges.  A number of engineers
> know just enough to be dangerous and suggested the
> technique is to just make a shift register and run
> the incoming edge through n flip flops. "Everyone knew" that
> was the thing to do.  NOT.  Actually, I only had to
> reclock 2 times, but the delay between the two registers
> had to be MUCH longer than the 31 ns period of the
> 32 MHz clock.  I characterized the flip flops I
> was using to make a probability curve of metastability
> events per year vs delay time using a special fixturing
> board.  I ended up choosing a time delay that extrapolated
> to something like 1 event in 10 years, just about "never".
> In those days, 32 MHz clocking was considered very high
> speed for the available TTL logic.
>
Today as we design these things, the same basic "cure" is used in order 
to handle meta-stability. We re-clock 2 times, but often the data can be 
hand-shaked over the clock domain border so only a few lines is needed 
in each direction to be double-clocked. As we build modern designs, 
clock domain borders is a pain, but a known pain and it just needs 
enough learning to handle it. Often we try to keep most of the logic in 
one or a few core domains. The basic problem is as valid as back then. 
Over the years the practical way to solve it has changed with available 
technologies. Even in modern FPGA times it can cause... interesting 
problems. All the Register-Transfer-Logic type of the design is well 
confined, but clock-domain borders always remain "interesting". 
Dual-port memories or FIFOs tolerating clock domains shifts only solve 
"so much" it turns out.

Even when operating synchronously madness can hit, for sure. Those 
experience include issues with Signal Integrity problems.

The real morale of these stories is that it actually takes a lot of 
effort to make digital designs... behave digitally and robust. It is 
however worth the effort.

Meta-stability is never fully "solved", but sufficiently surpressed. 
This is just as with bit-errors, which we never really get rid of, just 
move them out of sufficiently often to be annoying.

Cheers,
Magnus