🚀 go-pugleaf

We have some in the lab. They work as expected, but we may not have pushed
the rough edges. I haven't tried fibers yet.

I think we only ran the old 70MHz versions at about 60. That was the
crystal we had handy. We have a pair of the latest and greatest as of
several weeks ago.  They are still marked "SAMPLE", but work at full
speed. Maybe the offical production ones are on the shelves now.

What sort of error patterns are you seeing? The 4/5 encoding on the
serial link means that a "simple 1 bit error" probably turns into a
clump of 4 mashed bits, or 8 if the error crosses a nibble boundry.
(That's assuming you are running in 8 bit mode.)

If the fiber link is the problem, you should be able to see "funny" data
going into the recv TAXI and/or VLTNs on the reveiver output. Sync the
scope back on the transmit side and look at the eye pattern. I don't think
there are any specs in the data sheet, but it has to be pretty crisp. More
like a good digital signal rather than the sort of signal that a fancy
modem would be happy with.

AMD has a bunch of slides that talk about TAXI error rates. They claim 200
ft of RG-58A/U gets a byte error rate of 10^-10 at 70MHz. (I think this
was with the 70MHz parts.) This is for 2 hunks of coax in parallel to
carry the differential ECL signal.

You could try cutting out the fiber gizmos. A twisted pair should work for
several feet. Alternatively, try making the fiber path worse by adding
some attenuation, say by making the fibers longer. If you have to insert a
lot more fiber before anything changes, the fiber link is probably good
enough.