[Top] [All Lists]

Re: My thoughts on the MB design...

To: riscy@pyramid.com, paul@suite.sw.oz.au (Paul Antoine)
Subject: Re: My thoughts on the MB design...
From: Drew Eckhardt <drew@nagina.cs.Colorado.EDU>
Date: Tue, 29 Jun 1993 00:37:49 -0600
In-reply-to: Your message of "Tue, 29 Jun 1993 09:35:34 MDT." <199306282335.AA03311@suite.sw.oz.au>

    Hello people,
        My original thoughts included putting single-chip ethernet and
        SCSI ports on the board, as these would give 32 bit IO in the
        critical areas, though I worry at being able to get them in
        something other than PQFP.

IMHO, sticking SCSI / Ethernet on the local CPU bus is overkill
and then some.  We need 1M/sec for the ethernet, 5 or 10M/sec
(With external SCSI cables, I often see problems doing SCSI
faster than 8Mhz) at the most for SCSI.

If we clock a 16 bit bus at 10Mhz, 0 wait states, we have 20M/sec 
split between 5M/sec or 10M/sec SCSI and 1M/sec ethernet,
more than enough. 

Running the SCSI and ethernet and a lower speed lets us use slower,
cheaper chips that come in smaller non-SMT packages.
        The ISA bus could be used for floppy/serial/parallel, as these
        cards are plentiful and cheap, whereas a similar amount of 4
        or 6 layer motherboard space is expensive.

The cost of motherboard space is a valid consideration.  In comparison,
how much space do we loose to the 62 pin and 36 pin card edge connectors
needed for ISA cards?   I've got my 386 case open right now, and they're
about the size of 4 SIMM slots side by side and end to end.

Like one of the other posters sugested, I'd like to reraise
the issue of building a propriety I/O bus, say 16 bits wide 
since that's what your popular, affordable chips are, clocked
at some reasonable speed like 10Mhz (ie, CPU speed / 4).

I'd also like to recommend that we move the low-bandwidth I/O
off the system board (keeping it on a 10Mhz bus of the 3070
or whatever) to these slots, since 

1.  There isn't a real difference between running a trace to
        something on the mainboard and taking a chip select out
        to a connector.

2.  As people have stated, realestate on a 4-6 layer board is 
        expensive.  Since these smaller peripherials aren't
        in massive PGA packages requiring separate ground and
        power planes, and hideous numbers of traces, we can 
        get away with cheap double sided boards.

3.  Many of the people who want a MIPS board have different needs.  
        One user remarked gimme at least for serial ports for terminals.  
        Some may find a dual-homed machine useful while others are netless.  
        Offering the ethernet circuit, extra serial ports, etc. 
        would mean that users could mix and match.

        We'd have more people wanting a mainboard so we have less
        of a chance of missing the lots-of-100 pricing since people
        can get their whizzy MIPS board without coughing up for 
        the ether board they'll never use.

        There would probably be enough  people wanting feature X to do 
        the subboard they're interested in lots of 25 or 100 so they 
        aren't stuck with sample quantity pricing.

Arguable, this does the same things that an ISA bus does.  However,
I suggest that it could easily replace the ISA bus

- ISA needs two big, fat edge connectors on board - a 62 pin and 
        36 pin.  

        We can get away with something that takes far less
        realestate, and circuitry, like a 40 pin (ie IDE) or 
        50 pin connector (say +/- 5,12 V, ground, 16 data, 
        DMA REQ/ACK, read/write, wait state insertion, clk, with the 
        remainder going to address lines.

- Since we put the address decode on the main board, have a 
        reasonabel DMA chip (I assume that the DMA chip handles
        scatter/gather?  Motorola runs DMA through the MMU so
        this is handled automagically, PC's broke it so anything
        with performance needs bus mastering $$$) our "boards"
        will be much less complicated to fabricate, easier to
        program (Look at the SCSI code in the Linux kernel - 
        every driver has different programming to deal
        with their own unique brand of scatter/gathering 
        busmastering DMA) and cheaper than ISA boards carrying this
        burden with them (since we aren't looking to get rich
        off this, and the boards will be simpler).

        The ISA expansion bus could readily be made VESA local bus,

Nasty timing problems.  

        thus giving access to a very broad range of video cards, where
        the VLB versions give full 32 bit access and acceleration.

All of the cheap video boards I've seen (say this side of $300)
have used DRAM, and you aren't going to write squat to the video
when you're competing with the video serializers (45M/sec 
for 1Kx768 at 60 hz, the "minimum" workstation resolution. Multiply
by an appropriate factor if you're one of those winers who can't 
live with less than 72Hz or 1280x1024 :-) ).

We would use VRAM (Some one suggested MAC VRAM SIMMs), which 
you'll have no problems writing to at speeds approaching 
50M/sec since it is inherently dual ported.

With that sort of bandwidth, and more processing power than 
any S3 board or TIGA, you're going to have *no problems*
getting excellent video performance.

As far as getting the S3 to bitblt, etc, you can do the same
thing if you have a DMA chip on the motherboard (Do the IDT 
chips snoop?)

        As for the 34010: there is no way that we can compete on
        performance, price etc. for FAST video hardware with the
        mass-produced VGA/SVGA cards.

Since we've been told about the 34076 / NS RAMDAC / timing 
generator, I don't think the 34010 is the best route to 
go in terms of either cost or performance.

As far as comparing it to ISA / VESA / PC crap : 
We've timed Tseng ET4000 boards at 6M/sec, it's not too 
hard to compete with that sort of performance :-)  Please 
read the paper on "smart framebuffers". 

As far as price : 

We're all guilty of wild speculation on this. 

So, why doesn't everybody pick their favorite chip (I called the 
NS distributor about the 16552, and will do the same with
the NCR 53c90 series of SCSI chips), call their favorite
distributor, and get pricing in sample quantities and lots 
of 100?

        I have no great objection to a SIMPLE video interface such as
        the DMA-driven one Neil mentioned

I agree.  It's an elegant solution, that should give admirable 
performance at the same time.

         though I think most people
        have VGA/SVGA video cards they could (and would rather) use.

I have a Trident board, and quite frankly it sucks rocks in anything
but monochrome.  I've used TSENG ET4000's, and they don't stack
up to our i960 based Xterminals that use a dumb frame buffer.  I've
used midrange S3 boards, and they don't stack up.  I've  seen
high end S3 boards, but for $500 they damm well better perform!
        As mentioned above, the main way of doing this is through the
        ISA/VESA bus, as it gives us access to sound cards,
        floppy/serial/parallel cards and a myriad others.

Which means you're paying for DMA on each board, which means 
you are paying for the relestate it takes for four SIMM slots
for every ISA slot you put in it, etc.
        There's also nothing stopping you putting space for a 'native R3000
        local bus' connector for those that want to design their own
        I/O cards of course [or those addicted to speed :-]
Local bus makes lots of sense for video, but in a low-mid range 
system you just don't need the bandwidth for I/O, especially
with the added complexity involved in the design.


<Prev in Thread] Current Thread [Next in Thread>