Chip-o-rama (was Re: Pentium II processor)

[chuck@Yerkes.com]

Okay, this got long.  it went from responding to Jeremy's flame bait
(based on Computer Scientology) and became Chuck's Random Thoughts
and Facts on various CPUs.

Quoting Chris Cappuccio (chris@sefl.satelnet.org):
> On Fri, 18 Sep 1998, Jeremy Tregunna wrote:
> 
> > Your first mistake was buying a Pentium CPU with MMX... if you can, take
Uh, he said Pentium 2

> > it back and get a Celeron 300a, they're much better CPU's (with less cache
> > however). That's just my opinion, i've found Pentium 2's are crap, Pentium
> > Pro's are still the top Pentium CPU, and Celeron is the next nicest Intel
> > CPU... MMX (in my testing) has resulted in blowing up a CPU (took 4
> > minutes to blow up a P2 400 (i was using it as a server cpu)).
Let's see, a "server cpu" uses different registers?

> > Bottom line: If you can, take it back, and get a Celeron 300a or splurge
> > and get a Pentium Pro 200, other than that, I really can't tell you how to
> > fix your problem but my guess would be the MMX shit. P2's aren't really
> > built for unices (from what I've heard).
Well, they don't run the X-Windows opcodes ;)  But with a
soldering iron, an electron microscope and a full moon, you can
patch it into the microcode (be sure to have extra electrons on
hand).  See the web site for detailed instructions.

> Hunh?  Blow up a CPU?  What the hell are you talking about?

You know - the "mltdwn" register;  it's a secret.  Intel put it
in and only told MS about it.  If you don't set it to 0 within
an hour of startup the CPU pulls extra energy just to create
heat.

Okay, for real:
The PPro and P2 are better/stronger/faster than the Pentium.
The P2 is BASED on a PPro core (lets call it the 686).  The PPro
seems to be sort of an orphan, replaced by it's brother just months
after it's introduction.

The Celery^Hon:  "For consumers"  (eg. a crippled P2).  You
    remove the cache, you make it cheaper.  Slower too, but what
    do consumers know.

AMD K6/2.  Nice chip, actually, for an X86.  Much cheaper than
Intel, fairly competitive.  Why by a Celeron when you can get a
cheaper K6?

MMX in general.  The 586, er "Pentium(tm)" came out with MMX.
This fits with the Intel philosphy that the CPU should do
Everything.  The advantage for Unix users was a better on-chip
Cache.  Now, if only the computer designers would figure out
that a smarter I/O system and graphics copper (with on chip
windowing support ala the old TI 32xxx chips) worked.  The SGI
O2 runs a processor each for those; so you can run a video from
the CD to the graphics system and the main CPU _never_sees_
the activity.  What if all you disk requests when out to a little
smart I/O system and you the cpu never had to worry about it.
Okay, that's close to SCSI, but I want more delegation.

x86 in general.  Still based on 8088, which was derived from the
Z80.  Ever run CP/M programs on a 586/200.  WordStar screams (as
did I).  Thanks to IBM and MS, we (well, you, I don't have x86
here) are still stuck with an architecture that's doesn't have
opposable fingers.

MHz:  You can't compare MHz speed of dissimiliar chips.  I learned
this with the PC-XT (4.77MHz) being pounded by the 65c02 (2MHz).
A 68k chip typically needs far fewer clock cycles per instruction
than an Intel chip.  RISC chips get most things done in one clock
cycle (or at least used to).

Speed?  Well, raw math doesn't really help a lot in computers.
A bit, but it's not the whole game.  Micron (early) came out
with a bunch of Alpha computers with something like a 2Hz bus.
Whatever, once the instructions got to the CPU, they ran REALLY
fast.  But most of the time the chip was in "wait state" for
memory and other I/O. 

 Memory hasn't really gotten much faster than when we had 66MHz
chips.  That's why the UVM code speeds up the Alpha so much,
lots fewer RAM accesses for VM access.
 A solution to slow memory?  Static cache RAM.  Intel has up to
512k (ooh! aah! why not the 2M built into the alpha?)

I especially liked the ads touting that the Intel chips were
"designed for Windows," uh, please show me the opcodes and
changes that NT demanded - you know the ones that are useless
for Unix?  A graphics card with an equivalent of QuickDraw/Windows
I could see, but not a CPU.

Either way, the CPU is only part of the system.  A PPC750 in
a NuBUS Mac just won't help you a lot.  OTOH, a 125MHz PCI
buss would do wonders for a Pentium 233.
--------------------
That notwithstanding (and some of this will be corrected, I'm
sure; but it should lead to an intersting discussion):

The PowerPC 750 (heart of the MacG3's) roars: Fast math, fast
integer.  A RISC chip, at 300MHz, it will eat a 400MHz P2.  The
PPC AIX boxes demonstrate this best.  The Macs tend to have
slower busses (66 or 80MHz) and that tends to equalize things a
bit.  But the PPC's are also really interesting in that there
are some really interesting microcontrollers.  Designed for
cars, or handhelds or other embedded apps; they often put cool
little networking things and A/D & D/A stuff onto the chip.
Perhaps an OpenBSD ROM image would fit in there.

The Alpha is the fastest.  The 21164 comes with, I believe 1MB
cache for each: data/instruction (maybe 2MB?).  It also does
some twisted things with instruction pipelining and reordering.
One of the reasons that it's hard to write great compilers for
it.  Digital's C compiler still generates tighter code than
GCC/Alpha.  The issue here is do they scale?  Until recently,
you couldn't get more than 4 way MP with the alpha.   That seems
to be being addressed.  The Alpha's biggest foe has been
Digital.  Apple apparently approached DEC to use the Alpha for
their New Macs; DEC wasn't interested.  Thus was born PPC.  In
the meantime, a quad processor Alpha (633MHz) is fairly cheap
and runs Linux (and under NT it runs x86 code faster than most
x86's).

UltraSPARC:  Finally.  That's the word.  After years of Super-,
Hyper-, Mondo- SPARCs that didn't run much quicker than far
cheaper PCs, Sun addressed the basic problem of chips that were
too slow for Real Computing (tm) with the UltraSPARC.  64bit;
fairly quick.  Not as quick as Alpha, but the Alpha had the
burden on Digital behind them.  OTOH, DEC's hardware guys were
not beholden to the Software guys and info about the machines
had always been easier to get than about Suns (talk to Linux
porter David Miller about pulling teeth for SPARCStation info).
64Way SPARCs are available and use the former Cray SPARC
divisions' crossbar switching for really fast I/O.  The E4000
and E10000 are quick on I/O.  But Solaris makes up for that
extra speed :)

(Strong)ARM.  Very nice, very low wattage chip.  Designed, I
believe for handhelds and portables.  Dunno enough about it
first hand, but people I respect admire it a lot; therefore it
will likely die at Intel's hands.


The MIPS R10000.  Nice chip.  Scales too - 128 processor systems,
shipping NOW (for more than you'll ever earn).  Instruction
pipelining and reordering, etc, etc.  Worse than and better than
the alpha in many ways.  Irix does some interesting things
(similiar to Sun's plans with Solaris 2.8 or 3.0):  You can
dynamically partition a machine, depending on needs.  I know
Sun's plans better.  Basically, you will be able to say:
"For now, MachineA is 6 processors, Machine2 is the other 10"
"For MachineA: Sybase gets 2 processors min; more if available.
 httpd gets 1 min, everyone else fights for what's available."
  That's the plan anyway.  Irix does (now) the former, dunno
about the latter.  Of course SGI has Pledged to Merced and
that's crippled MIPS even more; but it's still one of the
better chips out there.


MERCED.  Oooh, aaah.  As long as It's Not Out Yet, it's the
perfect chip.  Runs instructions before you think of them.  By
running faster than light, you can use idle time from the past
for your current work.  Need to break DES?  Under 0.1 seconds!

  Okay, real specs are sort of out.  It's going to be a bitch to
get compilers to work on it.  Best of all, DEC, SGI, and Sun are
all developing separate C compilers for it.  Who's will be
best?  Who knows?  Hopefully the GCC folks will get leaked info
and be able to write something good, but it's a really complex
chip with more promise that will likely be met soon.  Don't
expect to see it in any volume soon. (and by then, we should
have the first GHz chips coming out of Compaqital and maybe Sun
and IBM).

The 6502.  Actually could be considered the first RISC chip.
All instructions ran in 1 cycle, only 56 instructions.  Yeah,
yeah, it's 8 bit, but the characters we type are only 8 bit, so
it's perfect for email and stuff.  The 64k RAM limit might be an
issue, but bank switching has worked for ProDOS and MS-DOS.
I'll bet if they came out with a 600MHz 6502, the applications
would be unimaginable.

My KayPro-2.  Works.  Portable (given a luggage cart).  Fast.
Well, faster than a Selectric.


Okay, I'm sure others will chip in (pun intended) on this.

Love ya all

chuck