Captain Napalm <spc_at_armigeron.com> wrote:
> Intel did. It's the 432. Neat architecture. Slow architecture.
Slow architecture results in commercial failure. But it was
incredibly neat. It could best be characterized as a VCISC; it had
a *much* more complex architecture than a VAX.
Part of the reason it was slow was that they were pushing the limits of
1980 technology to fabricate it, so they couldn't actually fit the
main processor (GDP, for General Data Processor) onto a single chip.
They had to split it into two chips, of which the 43201 was the instruction
decoder and microsequencer, and the 43202 was the execution unit.
Unfortunately due to packaging limitations the microinstruction bus between
the two chips was not wide enough to support single-cycle microinstruction
execution.
The 432 was a capability-based architecture. Objects were divided into
two portions, an access part and a data part. The access part could only
store access descriptors (pointers to other objects), while the data part
could store anything but access descriptors. This allowed efficient
implementation of Dijkstra's parallel garbage collection. Unfortunately
each part of the object is limited to 64 kilobytes.
Intel claimed that the 432 was designed for Ada. Actually, it was
designed before Ada, and it merely happened that Ada proved to be a
particularly suitable language for it. However, the first HLL for the
432 that Intel provided to customers was a variant of Rosetta Smalltalk
called OPL:
http://www.rosetta.com/TechHistory.html#opl-432
432 GDP instructions were of variable size, from 6 to 344 bits, depending
on the number of operands (0-3) and the addressing modes. Intel had an
assembler for the DECsystem-10, but never shipped it to customers. The
cross-development environment for Ada ran on the VAX.
The 432 GDP did not do an I/O operations. There was a 43203 IP (Interface
Processor), which was used in conjunction with an AP (Attached Processor),
typically an 8086. The IP wasn't really a processor in the conventional
sense of the word, in that it didn't execute stored programs. It executed
commands provided by the AP in order to control five memory windows that
allowed the AP to access objects in the 432 memory space.
The 432 supported 16 megabytes of physical memory, and an extremely large
virtual space. Intel designed a persistent object store as part of the
iMAX-432 operating system, however the persistent object store implementation
was never shipped to customers.
Other chips in the 432 set included the 43204 BIU (Bus Interface Unit) and
43205 MCU (Memory Control Unit). Using the complete chipset, it was easy
to design systems with full fault-tolerance. All of the chips were designed
to be usable in pairs to support FRC (functional redundancy checking), in
which one component was the master and one was the checker. The checker
performed exactly the same computations as the master, but instead of driving
bus signals, it compared its own results with those of the master, and
signalled an error if there was a mismatch. The master and checker roles
were exchanged on alternate bus cycles in order to insure that there were
no latent faults in the checker.
FRC alone isn't enough to build a fault-tolerant system, because if a failure
is detected, it is not known which of the two components failed. So pairs of
FRC modules could be operated in lock step for QMR (Quad Modular Redundancy).
When an FRC pair fails, the board is disabled and the other board is known
to have valid state. Under the direction of software, several recovery
methods were possible. The remaining module could continue operating without
a backup, in which case the system is no longer fault-tolerant. The module
could be shut down, in which case the system is still fault-tolerant but
has lower performance. Or the module could be paired with a spare module
reserved for that purpose. The hardware provided mechanism, but the
software determined the policy.
Although it is out of print, the most easily found reference on the 432
is _A_Programmer's_View_of_the_Intel_432_System_ by Elliot Organick.
Anyone still have any 432 docs, hardware, or software?
Cheers,
Eric
Received on Fri Sep 18 1998 - 16:13:48 BST