[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Lucent Technologies & Sun Microsystems



>> Please compare apples with apples.  If you're content with that kind of
> 'crippled' NC, then you'd be satisfied with a similarly 'crippled',
> super-low-cost PC.  But the bottom line is that people are NOT satisfied with
> such crippled PCs.

>Thats because you are thinking of them as PCs. I wouldn't dream of
running a circuit design algorithm on my NC, but I might use my NC to input the
design and display the results. THE IMPORTANT ATTRIBUTE is TRANSPARENT 
distribution of the computation. 

What you're talking about is a (relatively) dumb terminal.  Frankly, I think 
that's an incredibly stupid way to misuse the considerable processing resources 
that is highly affordable to have locally.  And personally I think that a PC 
could easily run a reasonable circuit design algorithm.  

The real problem is that NCs simply aren't hugely cheaper than a PC which does a 
whole lot more and being a whole lot more 'standard' and a whole lot more 
expandable.  And let's stop this fiction about a $200 handheld NC, because it's 
safe to assume that you'll not be using such a device to display your circuit 
designs...

>In theory, we could design a NC that was extremely small and yet appeared
to have the horsepower of a Cray, by designing the applications so that
only the absolutely essential part of the computation ran on the NC and
the rest ran on the Cray. 

The fact that Crays are so NON-cost-effective is a good part of why Cray is in 
such deep financial shit.  Obviously **they** have a vested interest to try to 
pump up demand for mainframes and super-computers... even when it makes **no** 
economic sense.  If you take the NC plus the price of the piece of the Cray you 
will use, you'll probably find that a PC dedicated to your problem would do it 
for a lot less money, and perhaps even faster end-to-end (since you don't have 
to share it).

>That's a terminal you say. 

Yes.  And a terminal onto a wasteful machine with a far greater cost-per-MIPS 
than even the most expensive, high-powered PCs.

>Right -- its a terminal when thats the right thing to be and a little computer 
when the computation can be handled locally, 

And that's part of the problem... the NC concept basically makes it impractical 
to do much of ANYTHING locally, both because there's not much in the way of 
storage, painfully slow communications to download databases for local 
processing, and because of the outrageously inefficient processing design based 
on the machine-independent interpreted Java system.  If you take a good look, 
nearly all the people pushing the NC are those people with vested interests in 
the big machine/central server arena... IBM, Oracle, Unix companies, etc. etc.  
And it's no mystery why, either... their margins on this big-machine crap is 
HUGE compared to the margins they can get in PCs, where there's so much more 
competition.  And contrary to their blithe stories and dramatic waving of hands 
about "open systems", these high-margin mainframe-think systems tend to be 
**highly** proprietary, using nonstandard parts, non-standard upgrade 
components, nonstandard peripherals, etc. etc., where you're TRULY stuck again 
in a [bad/old] single-vendor world when it comes time to upgrade (and of course, 
with upgrades coming in giant expensive (and hard-to-negotiate) leaps rather 
than replacing a few industry-standard parts for a few hundred dollars).

Some of us guys have been in this business long enough to have been there, done 
that, got the T-shirt (and saved it in the closet... hung next to my bellbottom 
pants, just in case *those* ever come back into style too).  :)

>...but it has filesystems mounted half-way around the world, if need be, and 
can run parts of its computation on a CPU in the network.

A PC can do all of the same things that the NC can do regarding communications 
and connectivity, PLUS providing an **exceedingly** cost-effective and 
significant computing resource, probably at a lower cost than the cost of a 
similar amount of computing power made available through some kind of remote 
centralized server.

>> >They won't compete with PCs, but my PC, even a laptop can't fit in my
> shirt pocket and run on a couple of AAA batteries.

>> The Atari palmtop runs (ran?) PC software and MS-DOS, and fits in a shirt 
pocket and runs off batteries.

>And where did the apps run -- on the palmtop. 

Right.  Where you could use them in an elevator, on an airplane, in the subway, 
in the wilds of subsaharan Africa while on a safari.  ANYWHERE.  You can walk 
into a client location and demo the product you're working on, without having to 
figure out some cockamamie way to hook up to their phone system somehow.

>I don't want my NC to be limited to the apps running on the palmtop -- 

A PC is no more limited to apps running on a portable than the NC is.  The 
difference is that the NC **depends** on running the applications (even 
downloading them on demand, which is PITIFUL) from another machine elsewhere.  
This kind of rubbish is far less justifiable today, now that notebook computers 
can easily be had with local storage of a gigabyte or more (and easily attach a 
highly portable Jaz drive or something for essentially unlimited removable 
storage.  The PC can run applications elsewhere IF NECESSARY, but isn't limited 
to that the way the NC is.  I think it pretty outrageous that your statement 
above tries to suggest that the PC is limited and the NC isn't, when the truth 
is exactly the opposite!!!

>they should run where they will run most effectively. Algorithm for determining 
same, TBD.

I have yet to see *any* large timeshared server computer which delivers 
computing performance more COST-effectively than a PC does.  If you can find 
one, please provide specific examples.

Note that in cases where a person needs *dramatic* amounts of processing power 
and locally, one perhaps very cool approach is to add a coprocessor board to the 
PC and stuff one (or several, for that matter) of the new TI DSP's on it... 
those DSPs deliver up to 1.6 Billion instructions per second per CPU, and 
that'll increase to 2 Billion instructions per second in short order.  And note 
that those DSPs in production quantities are less than $100 each!!  Many laptop 
computer docking stations (oftentimes still *quite* portable) have provision for 
one or two standard card slots, which would allow you to use such a coprocessor 
card while still staying surprisingly portable.  A small number of these TI DSPs 
will together deliver performance for many power-hungry applications that will 
compare quite favorably (and at orders of magnitude lower cost) than any kind of 
Cray or other big machine.

It simply *DOES NOT MAKE SENSE* to centralize CPU processing on a remote, 
centralized compute server anymore.  Or to have to download programs through 
relatively glacial-speed comm facilities before those programs can be executed. 
 At least not unless there are POWERFUL overriding considerations which justify 
the hugely increased cost of doing so.

Gordon Peterson
http://www.computek.net/public/gep2/