FPUs for small computers (Was: What to download for a PDP-8)
ajp166 at bellatlantic.net
Wed Oct 17 19:43:31 CDT 2007
>Subject: FPUs for small computers (Was: What to download for a PDP-8)
> From: Jim Battle <frustum at pacbell.net>
> Date: Wed, 17 Oct 2007 12:09:44 -0500
> To: General Discussion: On-Topic and Off-Topic Posts <cctalk at classiccmp.org>
>> Jules Richardson wrote:
>>> For a math copro to be called as such, does it have to integrate
>>> itself with the CPU's instruction set - or does some sort of box of
>>> tricks sitting on an I/O port and accessed through normal CPU
>>> instructions count? I'm not sure what the correct definition is.
>> It does tricks, but then with only 8 opcodes defined on the PDP-8
>> what do you expect.
>>> What do you mean by 'smallest computer' by the way? Smallest in terms
>>> of spec? I'm pretty sure some of the 1950's stuff had optional
>>> floating point hardware available, although it wouldn't have been
>>> physically small :)
>Northstar (of Northstar Horizon fame) sold an S-100 card that had a TTL
>FPU on it. It was a microcoded affair (256 words of 40 bits) that
>processed a nibble (BCD digit) per clock (4 MHz I believe). You could
>also specify how many digits were in the mantissa (2-14, even # digits
>only). Only the four basic functions, +-*/, were supported.
>Northstar had a version of BASIC that could use this FPU. As I recall,
>you had to specify how many digits of precision you wanted in your
>floating point numbers when you ordered BASIC (it wasn't dynamically
>specified in the interpreter).
>Others sold S-100 cards that used an AMD FPU chip. This too required
>poking data bytes and a command, then waiting for the result to be
>computed and then pulling out the result bytes. It wasn't somehow
>integrated into the instruction set of the host processor.
>Finally, I recall seeing an article where somebody took a pocket
>calculator chip and essentially poked simulated keystrokes at it and
>then decoded the LED driver output to determine the answer. It was very
>slow, though, so all it saved was the space of the floating point
Actually The article used a CALC chip that was slow on speed but had
an interface that was BCD output and convenient for input.
A few years back I took a 8742 (8048 with a slave interface)
at 11Mhz and programmed it to do 24bit BCD math (+-/*) and it
was faster than a calc chip. It was done because I could.
Overall it was as fast as z80/4mhz.
More information about the cctalk