Date   : Sun, 07 Aug 2011 23:15:38 +0100
From   : theom+news@... (Theo Markettos)
Subject: Risc PC (Was 'Minitel in France')

Tom Walker <tommowalker@...> wrote:
> > > I'll agree it wasn't a big deal in 1991, but it did fail to 'future
> > > proof' the machine in any way.
> >
> > MEMC was designed way before 1991...
> 
> But Acorn consciously decided to keep using it instead of developing a new
> chip. Along with VIDC. Both of which I think shortened the A5000s life.

The main reason for the 4MB limit on MEMC1(a) was the logical to physical
address translators.  There was only a limited size of L2P content
addressable memory (CAM) that could be fitted in whatever process was being
used (2 micron?  3 micron?).  So there's only 128 L2P registers.  MEMC1 was
pushing the limits with 32KB pages with 4MB RAM.

Note that ARM2/3 have no concept (IIRC) of physical memory: it's just
memory.  MEMC does the mapping between logical addresses (the lower 16MB)
and the physical DRAM that it alone talks to.  So any MEMC2 would have
needed to follow this model.  Furthermore, the 26 bit memory map limits the
amount of physical RAM to 16MB.  (MEMC hard-wired memory map is 32MB logical
RAM, 16MB physical RAM, 16MB I/O including ROMs)

I imagine A5000 development started sometime in 1990 (probably overlapping
with A540 development).  At this point ARM6 was probably under development:
ARM3 came out 1990 so Acorn's processor engineers would have started on the
next generation.  That next generation, which became ARM6, included a full
on-chip MMU including traditional page tables and TLB.  That radically
changes the memory architecture as the processor now knows about virtual and
physical addresses.

So, standing in 1990, you know that to-be-ARM6 is in the pipeline. Designing
and fabbing chips is slow and expensive.  Are you going to commission a
MEMC2 knowing that it'll be obsolete when ARM6 comes along?

Don't forget that A5000 was Acorn's new 'middle' machine (replaced A400/1
series).  The serious professional user had A540, which could take 16MB RAM
through pluggable MEMC boards.  And the cost of the RAM was more than the
whole machine, so if you needed 16MB RAM you could buy an A540 with the
small change.  The limitations of A540 with RISC OS 3.1 were DD floppies and
the lack of PC-style I/O world (IDE, 16450 serial, EPP parallel).  But SCSI
was the workstation standard in 1990 - IDE was a poor relation.  And a
maximum-4MB A5000 fitted into Acorn's then policy of limited expansion
(remember that Acorn only officially supported up to 2MB in A3000s?  Just
like A310s were only officially upgradable to 1MB RAM)

VIDC was probably a worse decision: really what they should have done was
taken an ISA video chip and adapted RISC OS to talk to various video
hardware.  But was any interesting ISA video hardware available in 1990? 
Was it plain VGA and no more? All the ColourCard etc chips came along in
1992-3.  And supporting TV modes/monitors would have been a headache: VIDC
is /very/ flexible, so we're probably back into custom silicon territory.
Perhaps VIDC3 (A500 used VIDC1, Archimedes used VIDC2) should have been
started in 1989.  Fixing the bus contention issue with dedicated RAM might
have helped, though again RAM was expensive.  But maybe the engineers were
tied up with ARM3 and then ARM6?

I may have to try and corner Steve Furber over a pint sometime and ask :)

Theo