allisonp_at_world.std.com wrote:
>
> > Using discrete transistors to build flop cells on that scale would be so
> > large that getting the timing to work would be a real pain. Core wouldn't be
> > bad at all; mid 70's core technology stuffed 32K * 20 in an 8" x 4"
> > area with no sweat; drivers and transceivers wouldn't double that.
>
> Ok now delete all teh integrate circuits like the sense amps and substitue
> a flipchip sized module for each read amp (note some of those chips were
> duals and quads). Then look at the packages of diodes and use individual
> diodes.
My point is that using contemporary discretes I don't need a flip chip's
worth of real estate to implement a read or sense amp. The problem posed
was to build the thing using descretes, not discretes of 1960's vintage.
Equivalent power density and beta is available in packages sometime two
orders of magnitudes smaller -- although for core drivers the current
density would force things to be physically larger...
>
> > I have trouble with the notion of the uart filling a 9 x 11 board given
> > that I'm holding one that's occupying 4 x 5 inches in SSI. Yeah, the shift
> > registers would take a bunch of space but I don't see it using anywhere
> > near the amount of real estate suggested.
>
> Try a TTY PDP-8 omibus card. Keeping in mind yes the uart part of that
> is only part of the card it still has tobe interfaced. The ttl only
> version is a full quad card. Again as transistors it will grow.
That's fair; the original statement that I responded to had to do with
the real estate used by the uart, not the uart and bus interface. Adding
in the bus interface for the 16 bit machine this is out of I'm looking
at about 50% more area. I'm not sure about the scaling to transistors;
this board is *anything* but dense; packages are on what looks like about
a one inch pitch. That's more than enough area to stuff the equivalent
discrete logic, assuming I use contemporary tools.
> Having had my head in the RCSRI PDP-12 and LINC-8 plus the PB-250 I have a
> better idea of real machines using nothing but transistor technology.
>
> Granted modern parts could really improve the packing density.
That's the point; aside from some high current applications (like core drivers)
I can get a tremendous increase in density by using autorouted multilayer
boards and SMD. If a 3' box fan positioned .5" from the board (quick, whose 70's
data sheet is that a reference to?) doesn't keep it cool we can always immerse
it in freon...
> > multilayer boards while our '60s compatriots were laying stuff out by hand
> > on light tables (been there, done that, never again).
>
> Yes, touch way to go.
Funny, I found my box of rapidograph pens the other day...
>
> > I suspect you could build a pdp-8 using contemporary layout tools and discrete
> > technology that, excluding the core stack, was an order of magnitude smaller.
> > Using '70s core technology you could get the stack somewhat smaller while
> > increasing the storage density by a factor of eight. The 8080 would without
> > a doubt be larger than the '8 and slower than the NMOS version of the chip;
> > it would also cost a fortune to build.
>
> Maybe. If you cheat and use 70s core tech (ICs) that's a big savings as
> the core planes were never the total system anyway. It's the
> drivers/decoders that grow huge!
I wasn't considering cheating. Decoding is strictly combinatorial, so that's not
tough. Driving and sense is much more of a bitch, and I probably can't get
away with surface mount devices because of the current density requirements.
Even so, I can build a similar function op amp using contemporary discretes in
a hell of a lot less space than using 60's technology -- just the reduction in
size of the passives gets me a hugh increase in density.
> Also you'll never get the power/cooling of a say straight-8 even if size
> reduced to come down near as fast and the ultimate problem may be cooling
> or deliver the AMPs of power.
Power distribution would be an issue, but it's certainly a tractable problem.
I've lost context on the dissapation/cooling issue. Certainly we'd burn less
and dissipate less (although more per square inch) than the original 8 did
if we built an 8 using contemporary discretes, and a discrete implementation of
an 8080 would run much hotter than the original 8.
Ugh. Back to working in the 90's.
Best,
Chris
--
Chris Kennedy
chris_at_mainecoon.com
http://www.mainecoon.com
PGP fingerprint: 4E99 10B6 7253 B048 6685 6CBC 55E1 20A3 108D AB97
Received on Wed Oct 20 1999 - 11:36:23 BST