Intel strikes back with a parallel x86 design

[George Macdonald]

Not at all - I knew that. Spice was one of the first 'performance'
applications to make good use of the 80386. But it was and is
completely atypical of commercial software - e.g. it was run on
dedicated systems and could (and did) run without any security or
even an operating system (on some systems). And the occasional
crash wasn't a major issue, either.

I dunno what you consider "commercial software" but I already told you that
the 80386/87-16MHz versions of our software blew the 68020/68882-20MHz
away. Now maybe you don't consider Linear Programming "commercial" but it
is fairly numerically intensive and gives any CPU a good workout... IOW
it's a good measure of overall CPU capability & performance in general.
Seems, according to David, like the err, shoe fits.

At that time, you made a decent development environment available at
a reasonable price, and they beat a path to your door. With the
PowerPC project, they did - they were hammering on IBM's door for the
whole period I am referring to demanding that IBM release the thing!

I can only speak from my own experience again here - we had *paid* a
substantial sum of money to IBM for Risc/6000 systems -- small discounts
were not a sign of IBM's strong commitment, but nevertheless... -- which
went absolutely nowhere - the customers did not care... they wanted PC. We
were not about to dump more money into PowerPC along with all that already
dumped... on top of the wasted sums on previous minicomputer systems. We
were also not prepared to budget time to supporting a platform which had no
client interest and where performance advantage -- yes even Risc/6K -- was
barely noticeable.

Oh yeah, we did it one more time, with Alpha, when some idiot we employed
assured us he had clients lined up... who seemed to evaporate when it came
time to umm, commit. Never again - it's over and the Risc/6K should have
been enough to convince us. Nope we were not in a mood to "beat a path" to
anybody's door - it would have required a hefty arm-twist, free
hardware/support and a lot of $$ to get us to pursue anything but PC by
that time.

Of course this legacy had a strong influence on our attitude towards
Itanium - could be one of the reasons for its failure assuming we were not
alone. Hell even PTC made an official announcement of err,
"de-certification"
http://www.ptc.com/partners/hardware/current/itanium_letter.htm. I guess
it took them a bit longer but they eventually got the message.. and have
even now "certified" AMD64 as a platform.

 

[Nick Maclaren]

I dunno what you consider "commercial software" but I already told you that
the 80386/87-16MHz versions of our software blew the 68020/68882-20MHz
away. Now maybe you don't consider Linear Programming "commercial" but it
is fairly numerically intensive and gives any CPU a good workout... IOW
it's a good measure of overall CPU capability & performance in general.
Seems, according to David, like the err, shoe fits.

Sigh. The distinction I was using is the one that was used at the
time by IBM, Intel and others - and to a great extent still is.

"Business/commercial" (or "commercial", for short) includes everything
that is used in most offices etc. The fact that a lot of that was used
in academia longer before and still is (e.g. Email, text processing)
seems to have escaped the marketdroids and bean counters.

"Scientific/technical" includes most conventional programming, as well
as applications like Spice. It was and is regarded as much less
important by the marketdroids and bean counters, partly because it was
and is a much smaller market.

The fact of the matter (whether you like it or not) is that Intel
established itself as the chip maker for the IBM PC, which was never
intended by IBM to be used as much more than a programmable terminal.
Intel and Microsoft were not so blinkered, and realised the wider
potentials. However, in the days you are talking about, they were
targetting the small business/commercial market and trying to break
into the medium business/commercial market. While they would SELL
to the scientific/technical, they didn't regard it as worth changing
any plans for. And, to a large extent, they still don't.


Regards,
Nick Maclaren.

 

[Klaus Fehrle]

They do indeed plan to follow AMD's lead in this area as well. Intel
calls their implementation "Common Serial Interconnect", and it ties
in with Intel's plan to use a common system interface for both Itanium
and Xeon chips:

http://www.xbitlabs.com/news/cpu/display/20050615232538.html

Unfortunately we're going to have to wait until about 2007/08
timeframe before this actually happens.

Yes, I know about that one. However, the New Microarchitecture is
slated for late 2006, so predates that.[/QUOTE]

Assuming you refer to the hoax Merom and Conroe would be a new µarch,
don't believe the hype from blowing issue-width smoke. It's just higher
feature-yields of the very same silicon Yonah is packaged from initially.

I am somewhat puzzled at

exactly where each change is likely to be introduced, and how.

Again, the paragigm "Marketing is cheap, Silicon can be fairly
expensive" is a good starting point to understand Intels roadmaps.
Reading over recent papers on nowaday's design for manufacturabiliy
principles is even better though. Highly recommended. I gained
considerable drinking-time I don't have to use for puzzling roadmaps
anymore by means of this.

It is, of course, possible that Intel's engineers are, too, and
are running round in circles attempting to make reality out of
executive decisions.

I suggest to assume these chaps know exactly what they do, and why they
do it - but certainly not implying everybody at Intel is happy with
design-targets skewed to silicon as cheap as possible.

K.

 

[Ken Hagan]

Really? "Software" is still the key and there isn't a whole
lot for "Linux on non-x86". It's just a recompile just doesn't
impress the ISVs.

Quite. Ask yourself how many Windows apps are sold with an explicit
promise that they will run on WINE. That ought to be "not even a
recompile", and it still isn't done.

The majority of programs running on PCs are badly written by people
who no longer have a financial interest in fixing bugs or porting.
Either they've gone bust, or there is more money to be made release
new crap. Win64 is a large enough leap that most applications won't
make it. The same goes for .NET, much to Microsoft's annoyance.

These programs are run by customers who insist on being able to use
the lowest powered PC in their posession. Even now, many developers
have to spend time making sure their product "degrades gracefully"
on Win98, a distinctly non-trivial task.

 

[Nick Maclaren]

|> >
|> Assuming you refer to the hoax Merom and Conroe would be a new µarch,
|> don't believe the hype from blowing issue-width smoke. It's just higher
|> feature-yields of the very same silicon Yonah is packaged from initially.

Hmm. I think that you are slightly over-stating that, as Otellini
has stated publicly that they are going to support EM64T (and, I am
pretty sure, SSE3). That makes them at least a significant revision,
even if Yonah has those features in disabled mode.

I agree that the Merom, Conroe and Woodcrest core is just the Yonah
core on steroids. Still, that sounds like a fairly decent design.

What is massively unclear is whether they are going to be the
Yonah core with a hacked, interim Common Serial Interconnect
FSB replacement, or that they are going to be purely FSB-based.
I have heard both rumours from fairly reliable sources.

|> I am somewhat puzzled at
|> > exactly where each change is likely to be introduced, and how.
|>
|> Again, the paragigm "Marketing is cheap, Silicon can be fairly
|> expensive" is a good starting point to understand Intels roadmaps.
|> Reading over recent papers on nowaday's design for manufacturabiliy
|> principles is even better though. Highly recommended. I gained
|> considerable drinking-time I don't have to use for puzzling roadmaps
|> anymore by means of this.

Unfortunately, I need the answers for procurement. Tempting as it
is to spend that money on beer, the bean-counters might make some
disapproving noises.

I have a fairly good idea of what is possible here, but am not
sure how Intel are going to schedule it. What I am likely to get
are some claims of roadmaps (almost certainly not from Intel itself),
and I need to be able to classify them into:

That's OK, but might slip a little

Promises, promises - let's reserve judgement

That is an attempt to hoodwink us - let's can that one


Regards,
Nick Maclaren.

 

[Anne & Lynn Wheeler]

By 1989 (2 years before IBM delivered PowerPC), x86 was FIRMLY
entrenched in the software market.

more like 2 years before starting somerset and the design of power/pc

previous post giving x86 chip shipments in the late 80s thru sept.
1990 ... I'm not actually positive whether the numbers given for 1990
where just for up thru Sept (when the numbers were compiled) or
includes projected for ye1990
http://www.garlic.com/~lynn/2005q.html#35

RIOS followed ROMP. it was same/original 801 design .. modulo having
to add hardware protection domains ... since the original from the 70s
was single domain and also inverted tables with 16 segment registers
for virtual memory. at advanced adtech conference in the 70s, i had
brought up the fact that it had provisions for limited virtual memory
objects (using 16 segment register implementation). there replay was
that there were no hardware protection domains, that application code
would be able to execute inline segment register change operations as
easily as it could change general purpose registers (any protection
being provided by pl.8 compiler/libraries only allowing correct code
and cp.r only loaded validly generated pl.8 code for execution). rios
also provided no cache consistency ... not between I & D caches and
not for any smp configurations ... that is one reason my wife and i
resorted to the ha/cmp design point for being able to do rios scaleup.

I have paperweight on my desk that was given out in Austin to
commemorate RIOS announce ... which has title:

IBM AWD Austin - GTD Burlington
POWER Architecture

and below the chips: 150 million ops, 60 million flops, 7 million
transisters).

the executive my wife and I reported to when we were doing ha/cmp
http://www.garlic.com/~lynn/subtopic.html#hacmp

then moved over to headup somerset ... the ibm, motorola, apple, et al
.... effort. i somewhat characterized it as being able to add 88k cache
consistency and smp memory bus to a 801 design. this became 6xx, et
all ... aka powerpc.

misc. past 801, romp, rios, fort knox, somerset, etc posts
http://www.garlic.com/~lynn/subtopic.html#801

the following from
http://applemuseum.bott.org/sections/ppc.html

The first PowerPC processors were to be designed at Somerset and would
arrive at Apple in one year and Apple would introduce the
PowerPC-based Power PowerMacintosh series on the Mac's 10th
anniversary, January 24, 1994.

... snip ..

and
http://www.answers.com/topic/powerpc

from above:
Word
PowerPC Year Size # of Macintosh
Model Intro. (bits) Trans. Models

970 2003 64 52M G5
7400 1999 32 10.5M G4
750 1997 32 6.4M G3
740 1997 32 6.4M G3
604e 1996 32 5.1M
603e 1996 32 2.6M
603 1995 32 1.6M
604 1995 32 3.6M
602 1995 32 1M
601 1993 32 2.8M

... snip ..

here is another early perspective from bull's viewpoint
http://febcm.club.fr/english/power_pc_products.htm

it mentions rochester decided to not go with 620 ... but instead
of going their own way.

I've frequently described 801 (simple and KISS) was a reaction to the
horribly complex FS project ... which was conceled before it was ever
announced
http://www.garlic.com/~lynn/subtopic.html#futuresys

the folklore was then that some number of the FS diehards went off to
rochester and did s/38. fort knox era in the circa 1980 would have
moved their product onto an 801 processor (coming full circle).
however that strategy died also. s/38 morphed into as/400 with cisc
processor ... but as/400 finally did move to 801 processor with the
630.

this description somewhat jumbles story about use of "POWER" ...
since POWER shows up from the earliest period for RIOS (i.e.
including above mentioned paperweight). powerpc was to differentiate
that somerset/6xx stuff had PC design point ... not high performance
numerical intensive workstation
http://en.wikipedia.org/wiki/PowerPC

 

[YKhan]

Remember, AMD made even the lowly 8088 (the D8088, IIRC). They *were* a
second source for even the first generation x86 processors.

Yeah, but at that time AMD was cooperating with Intel. My assumption is
if you're cooperating, then you're not competing. So they likely just
followed Intel's lead in pricing there. Nothing like hating someone's
guts to start a price war (or any war).

Yousuf Khan

 

[Del Cecchi]

I like the analogy, but I feel that both the 4004 and the Model-T
had an engineering coherence that the ROMP/RIOS seriously lacked!


Regards,
Nick Maclaren.

Romp and rios were two different things as I recall. Although they say
memory is second thing to go. As for Romp, what do you expect from a
processor designed in Yorktown.

 

[Anne & Lynn Wheeler]

Romp and rios were two different things as I recall. Although they say
memory is second thing to go. As for Romp, what do you expect from a
processor designed in Yorktown.

romp was 16bit processor that was supposed to be for the displaywriter
follow-on ... it was only after the project got killed ... that
somebody notice that you could port unix to any chip and call it
a unix workstation ... previous post
http://www.garlic.com/~lynn/2005q.html#38 Intel strikes back wtih a parallel x86 design

they had hired the group that had done the AT&T port to the ibm/pc for
pc/ix ... to do one for (office product division displaywriter) romp.

 

[Stephen Fuld]

Sigh. The distinction I was using is the one that was used at the
time by IBM, Intel and others - and to a great extent still is.

I absolutely agree with the thrust of your post. I am sure you remember,
but others may not know, that the IBM S/360 came as a base model and then
had "commercial" extensions options (mostly support for decimal arithmetic
and related stuff), and "scientific" extensions (mostly binary floating
point) so you could tailor your machine to your needs. But I have some minor
nits to pick.

"Business/commercial" (or "commercial", for short) includes everything
that is used in most offices etc. The fact that a lot of that was used
in academia longer before and still is (e.g. Email, text processing)
seems to have escaped the marketdroids and bean counters.

Most of the marketing people I know would agree that things like e-mail and
text processing originated and are still used in academia (and, for that
matter, in scientific organizations), but regard academia as a relativly
small market segment who seem to perpetually have no money and always want
steep discounts. Therefore, while they know what you said, they don't much
care.

Also, you seem not to mention what is truely the largest component of
"commercial workloads", the systems that actually run the businesses. These
are systems that keep the accounts for banks and insurance companies, that
handle sales and inventory for retail stores and distributers, make
reservations for airlines, rental car companies and hotels, handle pretty
much everyones payroll. etc.

"Scientific/technical" includes most conventional programming,

I disagree with the word "most". There are far more "commercial"
programmers than "scientific" ones, they write far more programs that
consume far more total CPU cycles.

as well
as applications like Spice. It was and is regarded as much less
important by the marketdroids and bean counters, partly because it was
and is a much smaller market.

The fact of the matter (whether you like it or not) is that Intel
established itself as the chip maker for the IBM PC, which was never
intended by IBM to be used as much more than a programmable terminal.
Intel and Microsoft were not so blinkered, and realised the wider
potentials.

Precisely true. An IBM bigwig (Akers?) was touring an IBM development lab
and saw his engineers using "personal computers" such as the old 8080 and
6502 based systems, asked what they were and why weren't they using IBM
computers. When he was told, he ordered a small project to create and IBM
equivalent quickly. The people in charge said the only way they could do it
quickly was to get a waiver from most of the IBM development requirements.
They got this "magic" letter that allowed them to go "outside" for many
components, including the CPU and the OS. The IBM executives thought it was
such a small market that the resulting loss to IBM was insignificant.

However, in the days you are talking about, they were
targetting the small business/commercial market and trying to break
into the medium business/commercial market. While they would SELL
to the scientific/technical, they didn't regard it as worth changing
any plans for. And, to a large extent, they still don't.

Agreed. And it probably was, and still is, a good business decision.

 

[Bill Davidsen]

IBM didn't own the minicomputer market in the early '80s. DEC did. IBM
captured a good chunk with the AS/400, but that was because the software
was better. There was no OS/400 running on x86.




...and IBM still sold *tons* of '386/'486 boxen. Go figure.




2000? Gee, I thought the '386 was out a tad before that.

Got a model number for an IBM rackmount server using the 386? We sure
never saw such a thing. You manage (as usual) to disparage without
provide a single verifyable fact.

 

[Bill Davidsen]

You really don't know anything about PC history do you? Applying the
term "low-margin" and "el-cheapo" to the 386 or 486 takes a leap that
you couldn't even make on the Moon. PCs cost between $2000 to $3500. At
that time, Intel still didn't have competition in the form of AMD and
then later Cyrix, so prices were kept high. AMD introduced the first
clones of 386 towards the end of the 386 era, and then later introduced
clone 486's. It's only after that point that prices started to plummet.
Until then, Intel was rolling in both volume *and* high-margins. We
never arrived at the era of sub-$1000 PCs until about five years ago.

I remember that when the 486 came out, it was a few months before the
dealer price for a system board with CPU dropped below $1000. And when
we started benchmarking them management wouldn't believe the numbers,
because they were faster than the mainframe.

Cyrix was not only a price competitor, their 387 was markedly faster
then Intel's at something we used, most probably transcendental
functions. For some programs the real time was about 2x faster with Cyrix.

 

[Anne & Lynn Wheeler]

Romp and rios were two different things as I recall. Although they say
memory is second thing to go. As for Romp, what do you expect from a
processor designed in Yorktown.

talk about fading memory ... and needing to dig something out of the
archives ... there is this nagging in the back of my mind about the
ibm 386 ... something about yields or something at intel (and meeting
demand) and ibm was going to fab some ... but ibm couldn't resist
tweaking the design. does anybody have recollection of what the chip
was called?

 

[Anne & Lynn Wheeler]

as an aside, i trip across refeence that a good 386 system was faster
than the (morphed displaywriter 16bit romp) pc/rt

prev. post with chip volume numbers from sept. 90
http://www.garlic.com/~lynn/2005q.html#35 Intel strikes back with a parallel x86 design

sjmn news article summary about x86 vis-a-vis 68k from april '89

Chip Wars:
1972 Intel 8008 3,500 transistors; 8-bit
1978 Intel 8086
1979 Intel 8088 29,000 transistors; 16-bit
Motorola 68000 68,000 transistors
1981 IBM picks Intel 8088
Apple picks Motoroloa 68000
1982 Intel 80286 130,000 transistors
1983 Motorola 68010 (adopted by Sun, H-P, Apollo)
1984 Motorola 68020 195,000 transistors 32-bit
1985 Intel 80386 32-bit; IBM adopts it for PS/2 line
1987 Motorola 68030
1988 Intel 80486 (1-million transistor mark broken)
Motorola unexpectedly releases advance details of 68040

Intel will take robe off 80486 contender today
Motorola is tying the gloves on the 68040

Time to watch some of America's most sophisticated technology companies
- slug it out over who will build brains of tomorrow's computers
- Microprocessor Wars, round 4
- contenders will find markets have undergone important changes
. 2-company contest has turned into a "bench-clearing brawl"

Drew Peck, Donaldson, Lufkin & Jenrette analyst, New York
- "Amount of confusion out there is as high as I've ever seen it."
- "Motorola and Intel are well-situated to become the
microprocessor standards of the future.
... it's a toss-up as to who's going to come out on top."

68040 and 80486 are very similar from technology point of view
- predictable extensions of familiar product lines
- promise whopping dose of industry's most sought-after commodity
. Speed (2-3 times as fast)
. approaching multi-million dollar mainframes
- First computers to use them expected in mid-1990
- extraordinary design feats
. 68040: 1.2 million transistors
. 80486: 1.1 million transistors
. 80386: 275,000 transistors (for comparison)
- High-level of integration
. Space for frequently used data and instructions
. Floating point math units

Chip customers are NOT similar
- Intel: adopted by IBM and Compaq
- Motorola: Apple, Sun, Hewlett-Packard
- 1988: Intel and Motorola neck-and-neck in sales of 32-bit processors
- Neither new chip is likely to give one company an advantage
- But computer markets are beginning to skew the race
- Intel may have trouble finding a home for the 486 in PC market
. Computer makers are only now absorbing the 386
. $1,000 price for 486 pushes computer to $15,000 range
. Computer makers may decide the extra power isn't worth it
. The applications for that power aren't there yet
("How much faster can my word processor run?")
- Some think Intel will begin poaching on Motorola territory
. technical workstations
. "... a chance to sway Motorola customers"
. "Problem is all that RISC stuff got there a year ago."
. "Not clear all that business hasn't been spoken for"
- Michael Slater, Microprocessor Report

Sequent Computer Systems, Beaverton, Oregon
- early Intel customer for 486
- plans to string the chips together into a mainframe

RISC
- Sun and Mips elbowed their way into the microprocessor market
- Some limitations:
. Not appropriate for all computing applications
- Quick speed fix of RISC has proved to be an irresistible attraction
. Put pressure on Motorola 68000 line
. "If we waited for Motorola, we'd still be waiting"
- Bill Keating, Sun director of technology marketing
- Motorola introduced it's own RISC processor
- Intel, well known RISC-detractor, rolled out its 80860
- Betting on both horses
. Forestall the RISC competitors
. Use faster conventional products to prevent further defections

Motorola is the company with the most to lose
- working overtime to smooth customer feathers
- gave reporters an inside peek at the 68040
. timed to coincide with Intel 486 announcement
- Souped up versions of the 68030 could, theoretically,
keep up with the RISC chips

Bill Keating, Sun
- 68040 and 80486 "are based on technology that's 10 years old"
- "They're serving the needs of software, more than hardware"
- How popular RISC becomes will prove the variable
. whether these chips represent a breakthru in technology
. or the end of the line

 

[Anne & Lynn Wheeler]

and another news article summary from a little earlier, sept. 88

* In a few months Intel will introduce its 80486 microprocessor
- mainframe on a chip
- 1 million transistors
- equal to low-end IBM 3090
* New age of personal mainframes
* Microprocessors will replace central processing units in mainframes
* Today Intel directly sells 20% of world's microprocessors
- 70% of all PCs use Intel technology
- 386 chip could be in up to half "IBM-type" PCs next year
* Intel's business is booming
- first half 1988 earnings up 211% to $224 million
- on revenue up 64% to $1.4 billion
* The 486 chip should sell initially in 1989 for about $1,500
- expect 486 based PC in 1990 for about $20,000
- such a PC should be able to run dozens of simultaneous large apps
- will greatly concentrate power into the executive suite
* Competing RISC designs promise ten-fold speed-ups
- Sun's Sparc chip is licensed by AT&T, Unisys and Britain's ICL
- MIPS Computer Systems RISC chip is licensed to Tandem Computers
- Sun and MIPS RISC chips will be produced and marketed by
TI, LSI Logic, Fujitsu, Cypress Semiconductors and others
- Motorola and Intel have introduced their own RISC chips
* Superfast chips of any design will open new opportunities in
application areas
- Intel safe in office applications thanks to $12 billion worth of
IBM-type PC programs but other areas are more open
* Mainframes could be built using these microprocessors
- standardized chips encourage mainframe cloning
- startups spared the up front costs of developing CPU and S/W
- Zenith Data Systems and Compaq Computer are planning to move
into minicomputers and mainframes using Intel's chips
- Intel already packages its chips into systems ranging from
printed-circuit boards to finished PCs
- Intel is in a 50-50 joint venture with Siemens to build a
fault-tolerant mainframe called Biin (buy-in?)
- Biin is due out this October
- Intel's systems business is 27% of total revenue this year and
is expected to be 33% in 1990
* Intel is improving its manufacturing capability
- closed 8 outdated plants
- spending $450 million this year on leading-edge equipment and
new plants
- let 25% of its workforce go, 6,000 workers and managers
- revenue per employee doubled last year to $100,000
* Parallel systems promise even better price/performance
- Intel has sold nearly 150 hypercubes containing 8 or more MPUs
- Hypercube can outperform a Cray at certain tasks
- Sequent Computer Systems of Beaverton, Oregon builds parallel
systems of mainframe performance that sells for one tenth the price
- Sequent's 120 MIP Symmetry system supports hundreds of simultaneous
users, putting it in the range of IBM's biggest 3090
- But Sequest's Symmetry costs $5,000 per MIP and IBM costs $120,000
- Intels biggest parallel system delivers 512 MIPs with 128 MPUs,
more than any commercial maimframe
- With 32 MPUs, Intel's IPSC machine outperforms a Crey X-MP/12
by 40% (100 megaflops vrs 70 megaflops)
- Cost per megaflop: Intel=$10,000; Cray=$100,000
- Others are selling parallel systems built from Motorola, Intel,
Inmos or National Semiconductor MPUs
- Boston's Stratus Computer, based on Motorola MPUs, has installed
1,444 systems at places such as Gillette, Visa International, and
the National Association of Securities Dealers
- Stratus' sales were $184 million in 1897 and have been compounding
at roughly 50% a year

 

[Nick Maclaren]

Also, you seem not to mention what is truely the largest component of
"commercial workloads", the systems that actually run the businesses. These
are systems that keep the accounts for banks and insurance companies, that
handle sales and inventory for retail stores and distributers, make
reservations for airlines, rental car companies and hotels, handle pretty
much everyones payroll. etc.

I didn't mention it, because it wasn't relevant. Back in the days
of the 80386, no serious company used an IBM PC for that! Intel's
second success was breaking into that market, but that came after
the PowerPC had failed.

I disagree with the word "most". There are far more "commercial"
programmers than "scientific" ones, they write far more programs that
consume far more total CPU cycles.

What those people do can't really be called conventional programming,
and quite a lot of the languages they use aren't even Turing complete
(ignoring finiteness restrictions). The conventional programming for
the "commercial" systems is done by a fairly small number of people
(e.g. the people who develop Oracle), and the vast number use those
higher-level programs.

I can witness that IBM used to regard the actual programming of even
some of the most "commercial" codes as a "scientific/technical"
activity


Regards,
Nick Maclaren.

 

[Del Cecchi]

talk about fading memory ... and needing to dig something out of the
archives ... there is this nagging in the back of my mind about the
ibm 386 ... something about yields or something at intel (and meeting
demand) and ibm was going to fab some ... but ibm couldn't resist
tweaking the design. does anybody have recollection of what the chip
was called?

was that blue lightning? I thought that was a 486, using the license we
got when we kept Intel from going bankrupt buying 10 percent of their
stock (newly issued).

 

[Nick Maclaren]

romp was 16bit processor that was supposed to be for the displaywriter
follow-on ... it was only after the project got killed ... that
somebody notice that you could port unix to any chip and call it
a unix workstation ... previous post
http://www.garlic.com/~lynn/2005q.html#38 Intel strikes back wtih a parallel x86 design

they had hired the group that had done the AT&T port to the ibm/pc for
pc/ix ... to do one for (office product division displaywriter) romp.

A question for all you omniscient ones out there - what were the
worst computers of all time? IBM's candidate must surely be the
PC/RT, but A,T&T are strong competitors with the 3B2.


Regards,
Nick Maclaren.

 

[Nick Maclaren]

and another news article summary from a little earlier, sept. 88

That's nice! Let's look at a few points and see how things really
developed.

* In a few months Intel will introduce its 80486 microprocessor
- mainframe on a chip
- equal to low-end IBM 3090
* New age of personal mainframes

Hmm ....

* Microprocessors will replace central processing units in mainframes

True. Just not in the 1980s ....

- expect 486 based PC in 1990 for about $20,000

What was someone saying about the PowerPC being unable to sell if it
were as much as $5-10,000?

- such a PC should be able to run dozens of simultaneous large apps

True. It should be able to

- will greatly concentrate power into the executive suite
!!!

* Competing RISC designs promise ten-fold speed-ups

They did, didn't they?

* Mainframes could be built using these microprocessors

For some meaning of "could", I suppose.

- startups spared the up front costs of developing CPU and S/W

Hmm ....

- Intel is in a 50-50 joint venture with Siemens to build a
fault-tolerant mainframe called Biin (buy-in?)
- Biin is due out this October

I missed that. It must have got Biinned.

* Parallel systems promise even better price/performance

They do, indeed they do. And they still do

- Sequent Computer Systems of Beaverton, Oregon builds parallel
systems of mainframe performance that sells for one tenth the price

It was most impressive, that one.

- Intels biggest parallel system delivers 512 MIPs with 128 MPUs,
more than any commercial maimframe

Nuts. The 3090/600 delivered 960 Mips - truly useless ones, as all
operations had to be register-register, but it got there ....


Regards,
Nick Maclaren.

 

[Bill Davidsen]

The Intanium may be a revolutionary design - but this design isn't
able to revolutionarize the market because of its ridiculous integer
-performance and its even more ridiculous value/price-relationship.
It's good for number-crunching because this can usually be parallel-
ized by the compiler; but nothing else.

I'm not sure what you're seeing, benchmarks for integer a reported to
have improved after the next compiler generation. Like any new CPU some
source code will cause generation of suboptimal code, and most compiler
tweaks to recognize these cases not only help the benchmark, but user
programs. Yes, there are exceptions to everything.

And I expect furure CPUs to to go a step back and completely drop
out-of-order and speculative execution and to implement massive SMT
like that of the Niagara and multi-core-designs.

I doubt that features which speed execution of linear and suboptimal
code will disappear. If you mean that pipelines will get shorter I
agree, but "disappear" would be a huge hit. I don't know about "massive"
SMT, just because it doesn't obviously scale all that well. Providing a
2nd thread to use CPU units otherwise idle, ala Intel HT, seems to
provide a modest gain in performance while requiring smaller gains in
cost, yeild, and chip space. More threads would probably starve for
execution units unless you added more of those, and complexity goes up.
I think Intel hit the sweet spot, measurable gains in performance
without much cost. There may be other sweet spots, but I'm not sure
they're at the mass market level.

As the number of cores goes up the bandwidth to memory better go up as
well, and that includes everything in the path including the memory
devices. Server memory has had 2 and 4 way interleave for a long time,
I'm guessing that width is easier to add than speed, and that will come
sooner than later. Plus there may be issues which need another
generation of fab to solve, regarding power usage per core.

I guess I agree with your direction, but I don't think the changes will
be as drastic as you predict.

I think you don't have access to the details of the process it
needed to change the P4-architecture to become AMD64-compatible.
So you can't say that definitely.

I read that Intel was developing an AMD compatibility chip about 18
months ago. I read it here, and many people said it was FUD.