From: John Gilmore <gnu@toad.com>
Date: Fri, 20 Sep 1996 18:48:54 +0800
To: cypherpunks@toad.com
Subject: Re: 56 kbps modems
Message-ID: <199609200744.AAA07457@toad.com>
MIME-Version: 1.0
Content-Type: text/plain


I checked the Rockwell home page, which has a pointer to the press
release.  It isn't very technical, but gives some good clues.  It
looks like they're doing an interesting trick.  The modems aren't
designed for use like traditional modems, where the same equipment is
on each end.  Instead, there is a digital interface on one side of the
phone call (like at an Internet Service Provider).  The consumer side
modem has a traditional analog interface.

The rest of this is speculation and fantasy on my part.

So, think about it.  The analog side will generate voltages and send
them to the local central office, where they will be digitized and
sent to the destination central office digitally.  There, they will be
patched into one channel of a T1 line (out of 24) and sent digitally
to the ISP's "modem bank".  Equipment is already available (and in use
all over the uunet network) that plugs T1 into a board full of digital
signal processors, decodes each of the 24 channels (each channel
running any modem signalling protocol, or ISDN), handles PPP packet
framing, and gateways the resulting packets to/from an Ethernet.

Now for Rockwell's trick, you get the DSP's in the two modems to talk
to each other.  They can run some simple coding scheme (say ordinary
2400 baud modem for this example) to pass digital data back and forth
while they're negotiating the full blown deal.  First, the analog side
sync's up with the clock for the 8000 samples/sec that the central
office is digitizing (into 8-bit samples).  You can do that by sending
one voltage and then switching to another; the far side can tell you
whether you switched on a sample-boundary or not (was there a sample
"in between" before it settled to the new value?).

OK, then, in each sample slot, the analog side can send one of 256
different voltages.  The digital side can tell it the 8-bit values it
received.  Then fine-tune that to sending 128 different voltages,
taking particular care around the ones that got distorted the first
time.  As long as you can find 128 distinct voltage levels that the
central office will reliably digitize, you're done.  You're sending
7-bit samples at 8000 samples/sec.  Do something similar for the
analog receive side, and you can start passing user data at 56K.

If the robbed-bit stuff gets in the way of seeing 128 distinct voltage
levels in every byte, you can send solid zeroes or solid ones in each
direction and see which bits they're stealing out of which bytes.
Use most of the 8 bits available in the other bytes (you can find e.g.
200 different voltage levels that will work), and in the stolen byte,
you can find e.g. 100 voltage levels that work.  This is more bits
than using 128 voltage levels in every byte, and in fact you can
probably get closer to 64 kbits/sec than to 56 kbits/sec, depending on
the analog qualities of the wire to your central office.

A nice trick!  It won't speed up analog-modem-to-analog-modem
connections, but those will increasingly be a smaller and smaller
fraction anyway, as the digital infrastructure becomes cheap.  And
of course the 56K modems will just be DSP's with decent A/D and D/A
interfaces, so they can run all the old analog protocols too, in the
case that the phone line isn't digitized, or if they want to talk to
an old modem.

	John