ph600fbj@sdcc14.ucsd.edu (james dehnert) (05/10/91)
It has occured to me after reading the post on how to put your 030 board in your 040 cube and make it run post that perhaps NeXT could make a multiprocessor board to boost preformance on the cubes. Imagine a board that you could put 1 to 4 68040 ( or even 88110 ) chips on right in next to your 030/040 motherboard. The Mach kernel can handel this ( can't it? ). I'm sure you would need quite a bit of memory to get it to work really well, but I'm convinced that MY cube isn't going to run the way I want it to until i get at least 40 meg of RAM ( not that I am unhappy now, but I do love speed ) So people of net land, what do you think? -- ******************************************************************************** jdehnert@ucsd.edu "The sky above the port was the color of television, tuned to a dead channel" -- William Gibson -- Neuromancer -- Eschew Obfuscation ********************************************************************************
dlw@Atherton.COM (David Williams) (05/11/91)
Instead of multiple 68040's or 88010's perhaps NeXT will/should go with HAL computers new cpus which are essentially next-generation of IBM's RIOS stuff. These babies will apparently blow the doors off of anything out there now (including HP's PA-RISC snakes) and Mips new stuff. This would give NeXT 2 advantages 1) Finally shut the mouths of the industry pundits who ignore NeXT for not using RISC. 2) Give NeXT the performance to hammer Sun/Mips and the rest of the workstation bunch for a long time to come. They'd now have a screaming box with a state of the art environment and apps. It would be way cool to just be able to slide one of these babies into a slot on the ol cube and just leave the rest of the workstation vendors in the dust. I'm sure Mr. Jobs would just love to have boxes that smoke the current and next gen ones from Sun, DEC, Hp and IBM. He'd even be burying Apple's RISC box before it left the starting gate. NeXT boxes using this technology could just become the "high end" of NeXT's product offerings in the $12k range...thought it would be nice to have this kind of power in the $5k range eh? David Williams
scott@erick.gac.edu (Scott Hess) (05/16/91)
In article <19229@sdcc6.ucsd.edu> ph600fbj@sdcc14.ucsd.edu (james dehnert) writes:
It has occured to me after reading the post on how to put your 030
board in your 040 cube and make it run post that perhaps NeXT could
make a multiprocessor board to boost preformance on the cubes.
Imagine a board that you could put 1 to 4 68040 ( or even 88110 )
chips on right in next to your 030/040 motherboard. The Mach kernel
can handel this ( can't it? ). I'm sure you would need quite a bit
of memory to get it to work really well, but I'm convinced that MY
cube isn't going to run the way I want it to until i get at least 40
meg of RAM ( not that I am unhappy now, but I do love speed )
I'd say get the extra memory.
The problem with doing a multi-processing cube is that in many cases,
it simply would not increase the speed. Without some fairly radical
restructuring of how the programs on the NeXT work, at least. What
parallel processing would do is allow people to do multiple disjoint
things at once.
_But_ the main speed complaints I've found people have with the
NeXT aren't related to how much they get done once they've started
doing it - it's how long it takes to stop doing one thing and
start doing another. This can be seen by going out and grabbing
either the Monitor program, or my TimeMon program. Let it run
for awhile while doing normal things, and you can see how much
the processor's being used (better with Monitor, I'll admit).
Basically, the processor is used a whole bunch when you switch
between programs, and not alot once you're in a program. For
instance, right now typing into a Stuart session, % CPU time being
used is lingering around about 10%. Similar results with Edit.
I think the main difference is that what is really needed is
a way to lower the _latency_ (time from when the user initiates
an action to when feedback occurs), rather than the bandwidth.
With the current structure of the NeXT (and Unix in general),
parallel processing gives mainly high-bandwidth. To restructure
things so that it can also give better latency will to some
extent cripple the performance on non-parallel machines.
Back to the memory. As I was talking to someone at the MN NeXT
user's group meeting, it was mentioned that they were a little
afraid of buying a NextStation, as they expected a RISC or
faster '040 offering to be coming up. I don't think this is so
much of a problem. Our experience with '040 upgrades is that
upgrading the computer does not improve perceived performance
so much as upgrading the memory. A 16M '030 system will often
feel almost as fast or faster than an 8M '040 system, especially
in a heavily NFS-mounted network like ours.
The reason, of course, is that the biggest thing people notice
is the time between starting actions and when they are finished.
For most actions on the NeXT, this doesn't require all that
much processing power - unless you need to bring the OS in and
swap some pages to the swapfile. On a 16M system, you can
reasonably run a couple NextStep programs almost entirely in
memory, while on an 8M system, you can barely run one.
I feel fairly strongly that I'd rather have a 16M or 32M 25Mhz
'040 than an 8M 40Mhz '040.
Lastly, some disclaimers. For those of you doing stuff like
image processing, or large Mathematica calculations, or speech
processing, or AI work - yes, the fast machine will make a
difference. That's because you've got lots of stuff to keep
it busy with. Same with parallel machines - you can find stuff
to keep the extra processors busy. What I'm aiming for is the
sort of generic user. These are users using the machine for
programming, or desk-top-publishing, word processing, etc.
(This would basically be myself, when you get down to it.)
For these people, more memory is the big constraint, a faster
processor will come second, and multiple processors will run
third. This will probably change as we come across more and
more ideas about how UIs should work, and how operating systems
are structured, but for now, we're a bit stuck.
Later,
--
scott hess scott@gac.edu
Independent NeXT Developer GAC Undergrad <almost out!>
<I still speak for nobody>madler@nntp-server.caltech.edu (Mark Adler) (05/16/91)
In article <SCOTT.91May15131023@erick.gac.edu> scott@erick.gac.edu (Scott Hess) writes: >I feel fairly strongly that I'd rather have a 16M or 32M 25Mhz >'040 than an 8M 40Mhz '040. > >Lastly, some disclaimers. For those of you doing stuff like >image processing, or large Mathematica calculations, or speech >processing, or AI work - yes, the fast machine will make a >difference. That's because you've got lots of stuff to keep If you're doing large Mathematica calculations, you had better get the extra memory *and* the faster processor. Mma is a major memory hog (at least 1.2 is), and you're in for a lot of wear and tear on your hard drive at 8M or even 16M. For Mma, a 25 MHz 040 with 32M will be a *lot* faster than a 1 GHz 040 with 8M, at least for the kind of stuff I was doing, since the processor would be waiting for pages to swap in. Mark Adler madler@pooh.caltech.edu