🚀 go-pugleaf

In article <1989Nov28.104128.8045@hellgate.utah.edu> michell@cs.utah.edu (Nick Michell) writes:
>I'm interested in GaAs, although I don't much about Prisma.
>If the failure was technology related, was that due to the use
>of parts from Gigabit - high speed, but also high power and only
>MSI density (not to mention the fact the Gigabit is in financial
>trouble)?  The leading start-up in GaAs seems to be Vitesse, which
>trades off some speed for lower power and higher density (currently,
>around 15,000 gates).
>
>On a related subject, DARPA has funded a number of GaAs RISC chips,
>which have been reported in various conferences and acedemic publications.
>This research has, at least so far, resulted in no commercial spin-offs.

DARPA has so far funded only development of what might be called "bare" CPUs
and SRAMs.  Several more part types are going to be needed (the McDonnell
Douglas part needed branch target cache, FPU, MMU/cache controller, operand
memory pipeline controller, and a glue chip called the system controller.  I
imagine the TI chip would need about the same).  While for development/demo
purposes one can use e.g. 10K ECL for the glue, cache RAMs, etc. (even then
it's not easy), if your "production" system is going to be mostly ECL it
might as well be all ECL (see your BIT or Motorola rep).  The big advantages
of GaAs for embedded military systems (speed/power product, military temp
range, radiation hardness) mostly disappear if the system isn't all GaAs and
CMOS (would anybody like to develop a Mil-spec BiCMOS RAM, please?).

>
>Is GaAs just not up to it?  Is the technology too immature still?
>It certainly appears to me that the current Vitesse technology would
>do fine for a Risc chip set.  It appears from recent EE times articles
>that at least Convex and Solbourne think select use of GaAs is worthwhile.
>Any comments, netland?
>
I think the _published_ Vitesse technology isn't quite there.  The current
McDonnell Douglas CPU is about 22K transistors, and that works out to be
about the bottom edge of what you need to do a workable integer execution
unit.  As it is, you give up a lot (maybe a GaAs Acorn?).  For an Enterprise-
(R6000) class machine it's not clear to me that GaAs really gives you that much
over ECL.  For a Cray-grade machine where speed-power product is critical, then
yes.

>/Nick Michell
> michell@cs.utah.edu