Index
Home
About
Blog
Subject: Re: Any opinions on SGI ?
From: mash@mash.engr.sgi.com (John R. Mashey)
Date: Oct 09 1995
Newsgroups: misc.invest.stocks,misc.invest
In article <45b812$hkc@unogate.unocal.com>,
stgprao@seismic.sugarland.unocal.COM (Richard Ottolini) writes:
|> SGI had about 75% of the workstations at this year's SIGGRAPH and Pentiums
|> with accelerators about 20%. The latter was about the power of an SGI
|> two years ago and had impressive capbilities. With the Intel CPU series
|> coming out with a new powerful chip every other year- P6, P7, etc.-
|> it is only a matter of time before they catach up to the workstation
|> companies.
Can you say more:
a) I don't know what "power" is .. hence it is hard to reply
in any useful way.
b) "an SGI": as noted earlier, there are plenty of things for which
there is a ~1000:1 difference from top to bottom.
c) Can you say more about the benchmarks/applications?
(Note: I have no doubt that today's PCs can do some things as well as
2-year-old SGI low-ends. Everybody knows that if one can buy a
commodity, and be satisfied with its features, distribution channel,
product-line, and support ... then they'll buy that, and there are plenty
of people who are; many of the people who buy SGI's aren't them.
Note: I saw an amusing article today that said PCs with forthcoming
graphics chips would have the power of a $10,000 SGI supercomputer ...
While I'm pleased that somebody thinks we sell a supercomputer for $10K,
[we don't - our supercomputers cost more than that :-)], it does illustrate
the confusion that abounds.
However, I may be biased, but it is not clear that they catch up
any time soon... For example, everyone in PC-land is thrilled with
the incredible performance leap going to PCI bus, and indeed, it is a leap.
Of course, it has 50% of the bandwidth of SGI's low-end a few years ago.
More relevant, as a long-term issue, one might ask why there is *still*
more than 1-2 car companies in the world?
a) Cars are a much more mature business.
b) Cars are *far less* differentiated than computers in terms of
performance, features, and nature of support. This may surprise
some people, but big companies, in particular, sometimes view
computers somewhat differently than do consumers.
1) Sometimes price/performance is almost irrelevant, what
counts is being faster than the competition.
2) Sometimes what counts is being a big enough customer for
some vendor that they actually get some input into the
next products.
3) Continuity counts: suppose you buy 100 workstations,
and they work, but the next release of the OS causes
them to break. What do you do?
A: yell a lot, and remind your workstation vendor
that you have other alternatives, and expect a
patch, and probably get one.
4) Suppose you buy a 100 PCs, with some cool new
graphics cards, and the next release of Windows ends
up breaking them. What do you do?
A: Yell a lot, and call Microsoft's 900-number, and the
graphics board vendor, or the PC vendor [since it sometimes
isn't at all clear where the issue is]. Remind them that
you have other alternatives :-) Of course, 100 units ...
is nonexistent, statistically, in this business model.
(None of this is theoretical; our household is computer-
diverse, with 4 Macs, 2 PCs, and an SGI Indy; my wife just
went thru Win95 conversion on one of the PCs, and it was
fairly exciting; I think she had some early add-in board
with some kind of chip flaw exposed by Win95; in any case,
she threatened the machine many times with OS/2 if Win95
didn't behave. It still crashes some, so I expect she may
buy a new board. Of course, many people had smooth
sailing putting it up.)
c) Anyway, as a computer systems vendor, I'd claim there are at least
2 obvious stable points in the market:
1) Build least-common-denominator, commodity systems, or in fact,
spend as little as possible on R&D, preferably $0, and concentrate
on distribution. Do not create new markets.
2) Concentrate on innovation, and looking for the next market
(even better, creating the next market), working with the kinds of
leading-edge customers who always want more, and designing things
with features that somebody will buy. (and, BTW, even at the lowest
level, CPU chips often have features that enable rather different
kinds of systems, difficult or impossible to build otherwise.
The *unstable* point is: build systems that are different from
the commodity, but not differentiated very much, and compete with
them directly on price, distribution channel, etc. Of course,
if a company thinks it is doing 2) but isn't ... it has trouble.
For what it's worth, SGI is clearly a category 2) company, and is
probably further away from the Intel/Windows/NT market than most
other workstation companies, in the sense that SGI doesn't try to
design low-end commerically-oriented desktop workstations.
It is difficult to explain some of this, since the issues are often down in
some technical nitty-gritty details. I'm working on a good analogy for
later posting that might help.
-john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc>
UUCP: mash@sgi.com
DDD: 415-390-3090 FAX: 415-967-8496
USPS: Silicon Graphics 6L-005, 2011 N. Shoreline Blvd, Mountain View, CA 94039-7311
Subject: Re: Any opinions on SGI ?
From: mash@mash.engr.sgi.com (John R. Mashey)
Date: Oct 10 1995
Newsgroups: misc.invest.stocks,misc.invest
In article <45ckr1$kjg@ixnews5.ix.netcom.com>, meyerkl@ix.netcom.com
(Kevin Meyer ) writes:
|> Agreed. There was also an interesting article in the San Jose Mercury
|> News today (10/9) about several companies coming out with multimedia
|> coprocessor chips. The gist of the article was that, with these chips
|> combined with Intel primary processors, the gap between workstations
|> and PC's will become small and probably not worth the extra bucks for a
|> workstation. All this while SGI is in the middle of a 3000 person
|> hiring binge... perhaps "pulling a DEC" does have some relevance?
That article says:
"could bring the multimedia ability of a $10,000 SGI supercomputer to
the desktop PC."
1) No numbers are given, of course.
2) Again, I'm pleased that somebody thinks that one of our $10K Indy's
is a supercomputer. I'm composing this on one, and it's very nice ...
but anyone who would call it a supercomputer ... words fail me.
3) For some reason, if somebody *says* they have put 3D graphics on a chip,
that seems enough for people to assume instant equality. Analogy:
My VW Beetle has 4 wheels & a rear aircooled engine ...
therefore it is a Porsche 911. [No disparagement of VWs is implied ...
VW uses many of our computers.]
One more time: there is a huge range of performance in 3D graphics;
there are things that people want to do that *we* can't do for
another 6 years, even in multi-$100K machines. Of course, at the low-end,
the bar goes up every year, and some things that PC's couldn't do last year
become possible ... but that fact hasn't changed in years.
By-the-way, somebody asked me about the graphics example & Windows problems ...
I ran across an amusing item in Sept 11, 1995 EE Times, catching up, p.8:
"When <Microsoft's> Display-X appears, it will stanch a potential problem
that's just come to light affecting some graphics and video-accelerator
cards under Windows 95. The problem involves cards designed to take
advantage of Microsoft's Display Control Interface (DCI). In Windows 3.1,
DCI handled advanced graphics functions such as rapidly moving pixels to
the frame buffer.
However, DCI is not present in Windows 95. Neither is DirectDraw, which is
the successor to DCI. Therefore, cards relying on acceleration enabled by
DCI are currently out in the cold, and DirectDraw isn't available to
take up the slack." Goes on to detail workarounds, somewhat.
So, a programming interface appears ... then disappears, and if you
depended on it, tough, you get to scramble around, or wait until next
version of Windows 95...
That is, I may speed up my Beetle by jamming a V8 in there ... but
nbobody told me it would burn up in a year :-)
--
-john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc>
UUCP: mash@sgi.com
DDD: 415-390-3090 FAX: 415-967-8496
USPS: Silicon Graphics 6L-005, 2011 N. Shoreline Blvd, Mountain View, CA 94039-7311
Subject: Re: Any opinions on SGI ?
From: mash@mash.engr.sgi.com (John R. Mashey)
Date: Nov 01 1995
Newsgroups: misc.invest.stocks,misc.invest
In article <47837u$geo@inews.sc.intel.com>, Herman So <hso> writes:
> DEC. SGI doesn't have the resources to continue developing it's own
> processor. Processor development is getting very costly, not only
> because of fab requirements, but because of design teams. This is
> very bad new for SGI in the
I believe that this is a *mis-statement of fact* that simply ignores
reality. I have posted previously on this here; let me try again.
1) MIPS (SGI) designs chips, including microprocessors.
These days, it costs *us* perhaps $30M-$40M over 3 years to do
a new high-end microprocessor, less for derivatives or lower-cost
chips. it may cost Intel more ... but I can't help that.
Design costs are *not* going up as fast as fab costs, you can only
fit so many engineers into one small chip.
2) MIPS has long-term partnerships with companies that include 2 of
the top 4 chip vendors (by revenue) in the world. They do process
development, build fabs, and manufacture the chips. Some of their
engineers are here with us. Fabs may cost $1B or so these days.
At least 2 of the *companies* who do this have total revenues substantially
more than, for example, Intel.
Note, for example, that the HP-Intel deal derives, NOT from HP being unable
to afford the design costs, but from not being a volume chip vendor,
and thus having difficulty affording the *fabs*.
Anyone who knows anything about the chip business understands these numbers.
3) Now, let's try some simple arithmetic:
a) Suppose the CPU design costs are as I've stated.
b) Suppose there are 3 CPUs in progress at once; it might cost
about $30M/year for CPU design.
c) SGI (by itself) is a $2.2B company heading toward a $3B company,
hence the CPU design cost is roughly 1% of the revenue.
(Of course, in 1994, there were $5B of MIPS-based computers sold,
so the overall cost is more like 0.5% of system revenue.)
d) The *most difficult* issue of having leading-edge CPU designs
is having good design teams, of which there are startling few in the
world; this is actually one of the biggest barriers to entry to do
a new CPU (along with software and giant CPU verification suites).
However, if you aready have built up teams and tools and software,
you can keep on doing it if the rest of the economics are OK.
4) So:
Is it worth spending 1% of your revenue to design CPU chips that
let you add value, differentiate your solutions in both hardware
and software? It may not be obvious to non-experts, but there are
features of chips that are *really important* to some people,
and those features have often taken much longer to find their way
into commodity micros. Some are not there yet.
Some people are working on kinds of machines
that absolutely require certain features not widely available,
nor required by the average PC user.
Actually, it isn't even 1%, because the MIPS-based computer business
was $5B in 1994, so it's more like 0.5% across that business.
Put another way, some things are justified on volume, others are
justified on revenue. On any CPUs *not* directly in the commodity
PC business, other costs far outweigh the CPU anyway.
Consider the following extreme example of volume & revenue:
Suppose you are a $1B company.
Would you invest $20M-$30M to design a chip that only sells in
10Ks, not in 100Ks or Millions?
Does that sound *really* stupid? Of course.
On the other hand, suppose that investment:
- Generates $300M-$500M of profitable, high-margin systems $$
- Gets you into a bunch of new sectors of computing, and
gets you in the door with numerous Fortune500 companies,
and helps you sell your other products as well.
- (and as a side-effect, helps New Zealand win America's Cup)
Maybe not so stupid after all ... (this was the R8000 chip;
SGI is now considered a major competitor in the supercomputer business,
where it had zero presence 3 years ago).
5) Finally, a slightly-exaggerated (but sometimes not) analogy:
VW makes many more cars than Boeing makes airplanes.
(Note: no disparagement of either company implied: both of them
are very good customers of ours :-)
How can Boeing possibly compete in transportation market
on such lower volumes?
Wouldn't it be better for Boeing to stop doing airplanes and do cars
instead? After all:
a) Many people own cars, go get groceries with them, and would find
no value in a faster car, or especially in buying an airplane.
They also buy individual cars, not fleets of cars/trucks/planes
that need to be compatible, have good upgrade paths, etc.
Nevertheless, by virtual of driving a car, they understand transport
issues fairly well :-)
b) It is easy to prove that cars are as fast as 747s:
race to the end of the runway: any fast car wins every time.
The price/performance of the car is much better at this.
(Note: certain graphics benchmarks are like this; I recall the Intel
i860's early demoes of Mandelbrots/sec, for example :-)
Hence, it may sound silly ... but we see it all of the time ...
somebody compares product A to product B, on work that both do,
but avoiding those things that B does, but A does not (i.e.,
ignore fact that 747 *flies*)....
(Note: I have seen otherwise sensible people say things like:
"a couple P6s plus a commodity graphics chip competes with a high-end
SGI workstation." ... and when somebody says this, they are acting
like a driver who thinks that a VW competes with a 747, perhaps never
having seen one of the latter, but clearing having *zero* clue about
differences there might be.
(A *high-end* SGI workstation is an Onyx or Power Onyx:
1) Uses real 64-bit CPU chips, up to 24/system. People parallelize
*real* applications onto multiple CPUs for performance, and the tools
are there to make that work fairly well.
2) Can support up to 16GB of main memory; certainly they've been
sold in the range, and SGI has shipped some related machines in
the 8GB-16GB range (I don't know if they had graphics or not).
3) Supports 64-bit operating systems.
4) Can support monstrous I/O
- Peak of 1.2 GB/sec of I/O
- At least one customer created a single disk file
377 Gigabytes long, and read or wrote it, sustained, at
300MB/sec, i.e., not disk caching :-)
- One of the common tests is to create a directory with
17M files - still works pretty well.
5) Supports up to 3 RealityEngine2 graphics units, which come with a
myriad of options, including such things as 48-bit RGBA
color with 192 color planes. [Huh? Why would anyone want
those things? Well ... why do pilots like so many fancy dials?
My car doesn't need that...] The *frame buffer* itself,
per head, goes up to *160MB* of memory...
Those things are multiple logic boards, each board with
more hardware than a PC motherboard ... and not because of
stupidity or lack of knowledge, but because it takes a lot
of hardware to do this stuff...
6) Is *not* typically used for word-processing or spreadsheets :-)
6) And has *far less* performance than many customers want.
I've been in meetings, with Fortune50 companies, where the
performance they want requires 2-3 more full generations
of hardware, i.e., after the year 2001, at least...
7) to do what these do, one needs CPU performance + compilers +
OS + graphics performance + I/O performance, and all at the
same time...
Of these characteristics
Somehow, I don't think 4 32-bit P6s + a 1-chip graphics unit, with a
132-MB/sec(peak) PCI bus ... shared by graphics ...
is the same thing ...
(Even SGI's lowest machines have a 267-MB/sec I/O bus...)
So: if somebody who *is* an airplane pilot wants to say that their race
car overlaps with low-end airplanes for some tasks (and they do), one might
have a sensible discussion. [In fact, high-end PCs and SGI low-ends do
overlap for some things. Nobody ignores that fact ... but the overlap
arises more from the fact that SGI keeps reducing the cost of its
low-end machines; SGI only started selling desktop machines in 1991.]
However, second-hand comments from non-pilots who heard that someone
else has seen a VW beat a 747, and therefore Boeing is out of business ...
do not help clarity of thinking very much.
-john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc>
UUCP: mash@sgi.com
DDD: 415-933-3090 FAX: 415-967-8496
USPS: Silicon Graphics 6L-005, 2011 N. Shoreline Blvd, Mountain View, CA 94039-7311
Subject: Re: SGI slipping
From: mash@mash.engr.sgi.com (John R. Mashey)
Date: Nov 08 1995
Newsgroups: misc.invest.stocks,comp.graphics
In article <aboucek-0511951706500001@aboucek.dialup.access.net>,
aboucek@panix.com (Alan Boucek) writes:
|> In article <47e0dv$216@b17news.b17a.ingr.com>, djallen@ingr.com (David J.
|> Allen) wrote:
|>
|> > But, let me ask this
|> > question: how good is your IRIX word processor? How about your spread
|> sheet, database,
|> > money manager, and other applications? Thats what I thought.
Wrong, people here use SoftPC/SoftWindows and run the same things...
of course, I doubt that many people are running money managers of the
kind you mean ... althugh we do sell fairly large machines to Wall
Street places for financial modeling, and Onyx's with 3D glasses to do
(whatever they do).
Note: this relates to the earlier discusion where someone asked
"What happens when people get enough CPU power, and PCs catch up?"
1) The question is an interesting one, but has some odd answers.
2) For some tasks, once the computer has gotten fast enough,
"enough IS enough", and making it faster only means that you run the idle
loop faster. In particular, in human-interactive programs, once the
response seems instantaneous (at high cost), you keep making it cheaper.
For example (already):
word-processors
calendar programs
many spreadsheet applications
presentations tools, 2D drawing programs
As examples of things for which enough might be enough, but nobody is
anywhere near close:
photographically (or 20/20-vision-accurate) out-the-window
driving simulator that is visually indistinguishable from
driving real car.
3) For some tasks, what matters is not so much how fast it is, but whether or
not it's faster/better than the competition, either directly, or indirectly
by its side-effects on goodness of design or time-to-market.
ex:
some trading software (the obvious one)
Some MCAD: the use of Pro/E, etc by New Zealand in the America's Cup.
Some high-end publish software, and some animation/special-effects
software (i.e., some people are always looking for cool effects)
4) For some tasks, enough is *never* enough.
Put another way, there are bunches of people who know *exactly* what they'd
do with more compute power than I ever expect to see... Their problems
*instantly* expand to fit new amounts of computation, memory, and I/O.
These often are people involved in modeling the real world, where denser
meshes, more atoms, closer approximations, more image resolution, etc ,etc
give better answers, and faster systems give faster answers, and more
solutions to be tried, etc.
ex: chemists, biologists, physicists, designers doing finite element codes,
computational fluid dynamics, etc, some of the oil-and-gas folks,
some of the medical imaging and medical modeling people, etc.
If this seems crazy, let me try an example: Paul Woodward, of the
University of Minnesota, and friends, wanted to run a 1024x1024x1024
fludi dynamics code, that is split a cube into a billion subcubes to model
fluid flow patterns in 3D. 1024^3 is not a particularly dense grid (in some
sense), but was necessary to get beyond computational artifacts found in
using only 512x512x512. If they could run the bigger problem, they could
do new science, and see effects invisible at the lower resolution.
Minor problem: there are 1B elements, and each element needs 6 32-bit
floats, or 24bytes/element, or 24GB of main memory. The only way to do that
at the time was to gang together a bunch of SMP servers (28GB), which they did,
generating 200GB of data for later visualization.
What would they need to get 2048^3 ... well that's 192GB of main memory,
and enough CPU power to do something with it.
Have various customers asked when we'd have such configurations?
A: Yes.
Q: Were they serious?
A: Yes.
Now, many of SGI's customers are in category 3 or 4...
5) Actually, if you believe that at some point, commodity CPUs will
be enough for most people (and that's likely true), who does that bother
more: Intel, or SGI? Hmmm. Consider the following question, addressed
somewhat by Wall Street Journal recently:
In the last round, there has been a curious inversion of business and
home purchases of PCs, that is, home buyers bought a lot of multi-media-loaded
Pentiums, whereas in earlier PC rounds, businesses bought the faster machines,
and home buyers got the less expensive machines, or waited longer.
The WSJ article speculated on whether (or not) the home market in the US
would start saturating in a few years, i.e., at the point where *home*
applications were perfectly well satisfied by Pentium CPUs, where
wealthier homes had several PCs/person, and less wealthy ones who could afford
any PC had one, and (the unfortunately large) number of homes that couldn't
afford one continued not to afford one. Likewise, perhaps the typical
business uses of PCs saturated, although maybe not.
IF you believe that (and who knows? may be true, may not be), then who runs
into the wall first:
a) people who sell mostly to home and commodity-business.
b) people who sell mostly to scientists, engineers, and creative
artists, at least some of whom think nothing is ever too much?
-john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc>
UUCP: mash@sgi.com
DDD: 415-933-3090 FAX: 415-967-8496
USPS: Silicon Graphics 6L-005, 2011 N. Shoreline Blvd, Mountain View, CA 94039-7311
Index
Home
About
Blog