88K Macintosh?

David Hough validgh!uunet!Eng.Sun.COM!dghaSun.COM
Mon Dec 24 12:40:26 PST 1990


If you want a RISC architecture that includes IEEE 80-bit extended precision,
your selection is somewhat limited, encompassing the Intel 960 family which is
positioned for embedded controllers, the WE 32206 which hasn't caught on,
and no others that I know about.  
[SPARC V8 and HP PA-RISC define identical 128-bit quad formats, 
so far unimplemented; as far as I know,
the MIPS book, the MC88100 book, and IBM America define only single and double.]

So the following provides an interesting possibility of a high-performance,
fully data-compatible future for Macintosh after the 68040, modulo data alignment
issues.  [Whether to support misaligned data is an interesting technical and business
problem.  For SPARC, Sun decided to make it fatal by default with an option
to just be very slow, in order to encourage people to reformat 68020 databases for
SPARC and obtain the best ultimate performance.    Hardware support for better-than-
very-slow misaligned access is a good thing during a transition but a bad thing
in the long run if it tends to perpetuate misaligned databases; then users 
blame the slow performance on the architecture or implementation rather than the
databases.]

The following comes from Unigram; I don't have any additional information:

            MOTOROLA PUTS ITS CASE FOR THE 88000
        BEING THE LONG-TERM WINNER IN THE RISC RACE
                   By John Abbott

Motorola Inc's Microprocessor Products Group has been revealing a little
more of its future product strategy in an attempt to quell the perception
that its late entry into the RISC marketplace has left it with little chance
of catching up with the currently dominant competitive offerings from Sun 
Microsystems Inc and MIPS Computer Systems Inc.  Details of Motorola's
second generation RISC, the 88110, first emerged at October's Microprocessor
Forum held in Burlingame, California, along with news of a new family of
embedded RISC processors, the 88300 family.  Since then, 88000 system
architect Keith Diefendorff has discussed the chip at the 88open Consortium's
General meeting in Frankfurt last month, and 88000 marketing manager Jeff Nutt
has headed a press tour through the US and Europe. 
 
Due out some time next 
year, the 88110 is a highly integrated version of its predecessors, combining
the functionality of the 88100 CPU and 88200 cache and memory management
units into a single chip performing at three to five times the speed.  Also
included on the new chip will be multiple integer, floating point and graphics
execution units to increase the parallelism within the chip - although just
how many units there will be is not yet revealed.  There will also be an
80-bit wide internal data path, improved handling of instructions sustaining
more than one instruction per clock, and branch acceleration to sustain
continuous execution.  The new chip will integrate around 1.4m transistors -
some RISC - and will run at a higher (currently unspecified) clock rate
than the current maximum of 33MHz.  

                           Sun's need

According to Jeff Nutt, Motorola's late entry into the RISC marketplace is
now turning to its advantage.  (Of course, had it satisfied Sun's need for
a RISC processor when first asked years ago, it would likely have eliminated
its major competition and continued to dominate the workstation market). 
Nutt claims that Motorola learned from the early RISC implementations, and
produced a flexible architecture designed to maximise the lifespan of the 
chip.  Thus Motorola will find it easier to add performance as others begin
to push their architectures to the limits.  Diefendorff says their are far 
too many players in an immature market (200,000 units shipped in 1989),
and that the fall-out will occur in 1992, with strong silicon manufacturers
emerging as the winners.  In order to extend its architecture to cope with
future requirements, Motorola has included room for up to eight Special
Function Units to provide for clean integration of additional functionality.  
 
In the 88100, a Special Function Unit was used to provide an integrated
floating point capability.  In the new chip, new functions and instructions
for graphics have been added in the same way.  Within the Special Function
Units, any number of multiple execution units can be added, to increase
parallelism.  Nutt describes the general approach of the 88000 as a
symmetrical superscalar, which reduces the load on the compiler and helps
with compatibility issues.  Even further into the future, Motorola hints
that the third generation 88000 will have more execution units, a wider data
path, and more parallelism, and will approach speeds of 100MHz, according to 
Motorola.  Remaining unconvinced by ECL, Motorola will implement the third
generation part in BiCMOS, the CMOS process that uses bipolar circuitry for
performance-critical functions. 

Assuming a quadrupling of performance by
each new generation, the 88000 family will be single chip, one million
transistor parts clocked at 300MHz by the late 1990s, and Diefendorff said
that there appeared to be no reasons why Motorola could not achieve a
throughput exceeding 4,000 MIPS by the year 2000.  Brave words.  Coupled
with the technology itself, Motorola hopes that the helping hand it has
given the 880pen Consortium in setting up a software support group for 88000
users will also help it win support.  880pen appears to have emerged as
the most successful software consortium to group round a particular
architecture: it popularised the whole area of binary and object compatibility
and applications binary interfaces among RISC vendors, encouraging AT&T Co's 
initiative to establish applications binary interfaces for Unix System V
release 4.  88open has a 500 page catalogue of software, the 88open
sourcebook, filled with what it says are all real, all available products. 
And it has steamed ahead with the most advanced conformance testing software
suites, which many other companies are now interested in licensing.  In
contrast, Sparc International has had a very confused history to date, and
until the liberation of Unix System V.4, Sparc users are closely tied in
to operating system, networking and user interface software that comes from 
Sun itself, although the licensing can be done through Interactive Systems
Corp.  Synthesis Software, the MIPS software initiative, ended in failure
because MIPS wanted it to make a profit, according to 88open president Tom
Mace, and even now MIPS has no application binary interface strategy in place.  

                              Mass860

And Mass860 group backing Intel Corp's 80860 currently appears to be a very
informal operation, with no clear mission.  But the group is in dire need of
a volume reseller.  Nutt claims that it takes time for such suppliers to come
out of the woodwork, and that they will duly appear over the next few years.  
Currently hardware manufacturers such as Bolt Beranek & Newman Inc, Cetia SA,
Data General Corp, Dolphin Server Technology A/S, Encore Computer Corp, Everex
Systems Inc, Motorola Computer Systems, Norsk Data A/S, Omron Corp, Philips NV, 
Quotron Systems Inc, Sanyo-Icon Inc and Unisys Corp have declared for the
88000, along with embedded control manufacturers, comms manufacturers, board
manufacturers and the military and avionics market, including Harris Corp
and Europe's Thomson-CSF SA.  Four hardware manufacturers still remain
committed but undeclared.  But the speculation remains about Apple Computer
Inc's RISC intentions, still not entirely clear - will it go for servers of
workstation products if it eventually does use the 88000?.  Compaq Computer
Corp and Ing C Olivetti & Co SpA have been mentioned as two majors that
have not so far announced their RISC intentions, but Motorola will face 
stiff competition for the business.  



More information about the Numeric-interest mailing list