🚀 go-pugleaf

solman@athena.mit.edu (Jason W Solinsky) writes:
>what are they getting in addition to computational power?

Simple: R&D, software, support, and lots of cool mugs, t-shirts,
highlighters, colorful posters, and slick glossies. -Pete

>solman@athena.mit.edu (Jason W Solinsky) writes:
>>what are they getting in addition to computational power?
In article <1992Nov15.200933.16043@news.cs.indiana.edu>
"Peter Beckman" <beckman@moose.cs.indiana.edu> writes:
>Simple: R&D, software, support, and lots of cool mugs, t-shirts,
>highlighters, colorful posters, and slick glossies. -Pete

How about >32-bit addressing: multi GB file systems, real 30+ MB/S file reads,
50 MB/S communication, decent numeric FP precision and performance, etc.

--eugene miya, NASA Ames Research Center, eugene@orville.nas.nasa.gov
  Associate Editor, Software and Publication Reviews
  Scientific Programming
  {uunet,mailrus,other gateways}!ames!eugene
Seeking Books to buy:	Bongard, Pattern Recognition
			3 down 1 to go.

In article <1992Nov15.201147.5302@athena.mit.edu> solman@athena.mit.edu (Jason W Solinsky) writes:
>It seems to me that supercomputing power costs about $250,000 per GFLOP
>nowadays. It also seems to me that a company could sell a GFLOP box that
>has enough memory, I/O and internal bandwidth to support near peak
>performance at under $10,000 and have PLENTY of profit margin.
>
>My question is, when people buy supercomputers at $250,000 per GFLOP, what are
>they getting in addition to computational power?

Amongst other things:

Support, libraries, a suitable OS, some apps, CPU cooling plant...

Damon
-- 
Damon Hart-Davis                          Internet: dhd@exnet.co.uk, d@hd.org

Public-access UNIX (Suns), news and mail for UK#5 per month.  FIRST MONTH FREE.
[1.35] Cheap Sun eqpt.  UUCP news/mail feeds.  Tel/Fax: +44 81 755 0077.

if you know how to build a 1GFLOP machine for $10,000
(much less sell it for that much), then you should do
so immediately and become immensely wealthy.

i don't know where you got this $10K number, but boy,
are you confused.

	-Mike

Bellcore?? Bellcore isn't allowed opinions - these be mine.

In article 5302@athena.mit.edu, solman@athena.mit.edu (Jason W Solinsky) writes:
> 
> My question is, when people buy supercomputers at $250,000 per GFLOP, what are
> they getting in addition to computational power?
> 

Memory bandwidth!

OK - maybe I was confused....

Barry Margolin suggested that maybe the original comment arrived
at $10/GFLOP, not a $10K 1-GFLOP machine.  That would mean a
$10M 1-TFLOP machine.

while i still think that's low, it isn't as unreasonable as
i first thought when i interpreted it as impying a $10K 1-GFLOP
machine.

i now return you to your regularly scheduled discussion.

	-Mike

In article <1992Nov16.023102.21232@nas.nasa.gov> eugene@wilbur.nas.nasa.gov (Eugene N. Miya) writes:
>>solman@athena.mit.edu (Jason W Solinsky) writes:
>>>what are they getting in addition to computational power?
>In article <1992Nov15.200933.16043@news.cs.indiana.edu>
>"Peter Beckman" <beckman@moose.cs.indiana.edu> writes:
>>Simple: R&D, software, support, and lots of cool mugs, t-shirts,
>>highlighters, colorful posters, and slick glossies. -Pete
>
>How about >32-bit addressing: multi GB file systems, real 30+ MB/S file reads,
>50 MB/S communication, decent numeric FP precision and performance, etc.
>

And several hundred gallons of Fluorinert(r) ;-).

Steve

In article <1992Nov16.141214.7632@walter.bellcore.com> mo@bellcore.com writes:
>if you know how to build a 1GFLOP machine for $10,000
>(much less sell it for that much), then you should do
>so immediately and become immensely wealthy.
>
>i don't know where you got this $10K number, but boy,
>are you confused.

Actually, even the $250,000/GFLOP number sounds a bit low to me; $2M/GFLOP
is closer to the current going rate.

In article <1992Nov16.174414.12427@craycos.com> jrbd@craycos.com (James Davies) writes:
>In article <1992Nov16.141214.7632@walter.bellcore.com> mo@bellcore.com writes:
>>if you know how to build a 1GFLOP machine for $10,000
>>(much less sell it for that much), then you should do
>>so immediately and become immensely wealthy.
>>
>>i don't know where you got this $10K number, but boy,
>>are you confused.
>
>Actually, even the $250,000/GFLOP number sounds a bit low to me; $2M/GFLOP
>is closer to the current going rate.

And for a *sustainable* gigaflop/sec, I'd say the number is even higher...

Steve

In <1992Nov16.171314.2178@craycos.com> sog@craycos.com (Steve Gombosi) writes:

>>>>what are they getting in addition to computational power?
>... several hundred gallons of Fluorinert(r) ;-).

Not necessarily. Some of us are ozone-friendly. :->

In article <1992Nov15.201147.5302@athena.mit.edu> solman@athena.mit.edu (Jason W Solinsky) writes:
>It seems to me that supercomputing power costs about $250,000 per GFLOP
>nowadays. It also seems to me that a company could sell a GFLOP box that
>has enough memory, I/O and internal bandwidth to support near peak
>performance at under $10,000 and have PLENTY of profit margin.

Ignoring for now the question of what counts as a GFLOP and whether a GFLOP
is really available in a realistic configuration for $250,000, and assuming
only peak rates matter, I still don't know of any design for a uniprocessor or
small number of processors in a shared memory configuration which as a parts
cost of $10,000 and can generate 1 GFLOP.  50 Megaflops maybe, 1 Gigaflop no,
not even close.  The current high performance processors (whether from HP,
DEC, IBM, or SUN) do not come cheap, and I am awre of none of them being
rated at over 200 Mflops for a single processor.  High performance memory
systems require expensive memory and glue logic.  My gut feel (without
calling any vendors for parts prices :-) is that the parts for a GFLOP
system would be at least in the $50,000 price range.  And that estimate does
not include any disks, tapes, etc.  Of course, prices are constantly
changing.  If you think such a design is possible, please post a rough parts
list and nominal parts costs.

Going from parts cost to full product takes on the order of a factor of 5
for low to moderate volume products.  Cost of assembly, including margin
testing, rework, shipping, inventory is not trivial when volumes do not
justify superautomation.  Then you need to cover other things, including
paying for R&D, sales and distribution, general overhead, and maybe even a
little profit.  Generally, if you skimp in one of these areas, your company
is nonfunctional.  Customers tend to not return to nonfunctional companies.
High volumes can help reduce markup, but supercomputers just aren't sold in
high volumes.  So if we start with a hypothetical $50,000 parts cost, we
see a shipping price in the range of $250,000.  Large companies can
selectively target markets for lower margins to gain market share, in the
hope that volumes will rise enough to justify the thinner margins.  A
startup company trying that strategy runs a big risk of going out of
business due to lack of cash before the strategy pays off.

It is also worth noting that any system that is shipping today started its
design 2 to 5 years ago.  A design started today should produce a product
which cost significantly less with a higher performance.  Unfortunately
it would not be available for a while, by which time others would do the
same thing.

In article <1992Nov16.174414.12427@craycos.com> jrbd@craycos.com (James Davies) writes:
>Actually, even the $250,000/GFLOP number sounds a bit low to me; $2M/GFLOP
>is closer to the current going rate.

I don't know what Cray's prices are, but $250K/GFLOP is the ballpark
figure for the CM-5. 

-- 
Ephraim Vishniac    ephraim@think.com   ThinkingCorp@applelink.apple.com
 Thinking Machines Corporation / 245 First Street / Cambridge, MA 02142
        One of the flaws in the anarchic bopper society was
        the ease with which such crazed rumors could spread.

In article <1e8tm6INN61f@early-bird.think.com>, ephraim@think.com (Ephraim Vishniac) writes:
|> In article <1992Nov16.174414.12427@craycos.com> jrbd@craycos.com (James Davies) writes:
|> >Actually, even the $250,000/GFLOP number sounds a bit low to me; $2M/GFLOP
|> >is closer to the current going rate.
|> 
|> I don't know what Cray's prices are, but $250K/GFLOP is the ballpark
|> figure for the CM-5. 
|> 
|> -- 
|> Ephraim Vishniac    ephraim@think.com   ThinkingCorp@applelink.apple.com
|>  Thinking Machines Corporation / 245 First Street / Cambridge, MA 02142
|>         One of the flaws in the anarchic bopper society was
|>         the ease with which such crazed rumors could spread.
Only if you mean peak.  And if you think a CM-5 can deliver $250K/GFLOP on
real application code, I have a bridge over the San Fancisco bay you might 
be interested in.

Sam

P.S., Don't feel so bad, our Cray can't deliver $250K/GFLOP either.  In fact,
we are lucky to get >1 GFLOP out of our Y-MP/8 on real code.

I am not going to bother trying to prove that $10K for 1 GFLOP peak is
very do-able. If you don't believe me that's fine, but I think I have
here the bits of information that I was looking for on why we don't see
machines with this price performance.

In article <1992Nov16.192316.22684@news.eng.convex.com>, Patrick F. McGehearty <patrick@convex.COM> writes:

|> Going from parts cost to full product takes on the order of a factor of 5
|> for low to moderate volume products.

|> It is also worth noting that any system that is shipping today started its
|> design 2 to 5 years ago.  A design started today should produce a product
|> which cost significantly less with a higher performance.

While I am working on a related project of my own, I originally asked the
question after noting the HUGE price difference between the low-end and
high-end AXP machines, and the comparatively low performance gain. Clearly
people are getting something for that extra money and its not superior
performance. I wanted to see what people who may have been involved in the
purchase of a supercomputer think they got for their money.

Jason W. Solinsky

What a silly question.

Blinking lights.  Lots and lots of blinking lights.   well, if one
buys a massively parallel supercomputer, one gets blinking lights.
From what I've seen of the C90, it doesn't have nearly enough lights.
At least it looks like a ST:TNG prop, I guess that's better than
nothing...

Founder, Society to Bring Back Lots of Blinking Lights


--
             J. Eric Townsend --  jet@nas.nasa.gov -- 415.604.4311
               '92 R100R, DoD# 0378, BMWOA, AMA, ACLU, EFF, CPSR
                                "HotHead Rules!"

Hardware performance monitors.

Not all machines have this, but boy it's handy when developing new algorithms.

Pete

--
Department of Computer Science, Duke University, Durham, NC 27706
Internet:	psu@cs.duke.edu
UUCP:		mcnc!duke!psu

In article <1992Nov15.201147.5302@athena.mit.edu> solman@athena.mit.edu (Jason W Solinsky) writes:
>It seems to me that supercomputing power costs about $250,000 per GFLOP
>nowadays. It also seems to me that a company could sell a GFLOP box that
>has enough memory, I/O and internal bandwidth to support near peak
>performance at under $10,000 and have PLENTY of profit margin.
>
>My question is, when people buy supercomputers at $250,000 per GFLOP, what are
>they getting in addition to computational power?
I/O, SCALAR performance, compilers, an operating system, folks who live
with the system and fix it when it breaks, the ability to run major third-
party codes and get the right answers, help with porting YOUR applications, and
a CREDIBLE growth path -- confidence that the vendor will be able to grow a 
system to meet your needs.

I'm sure in the next few years you will be able to buy a 1-GFLOP CHIP (i860 or
Alpha compatible) for a few hundred dollars; perhaps $10K for the board-level
version with some memory (will it be enough???) and some kind of I/O (is VME
fast enough????).  If cost per GFLOP is your criterion, you should be buying
an array processor board and putting it in a workstation.  But if you want to
do supercomputing, I'm afraid you're going to have to buy a supercomputer.

"patrick" == Patrick F. McGehearty <patrick@convex.COM> writes:
JET == jet@boxer.nas.nasa.gov (J. Eric Townsend) 

JET>Blinking lights.  Lots and lots of blinking lights....

patrick> have suggestions for specific hardware events that might be
patrick> connected to blinking lights on a massively parallel shared
patrick> memory super?

I sorta like the ipsc/delta/sigma solution:
 - one for cpu active
 - <something> for communication
 - one for node failure

I've seen more than one person debug their code on the ipsc/860 by
sitting in front of the cabinet and watching the led's.. "See, it hung
in computation after the third massive data shuffle..."

I'd also like to have some sort of cray-like debugging panel where I
could get information like:
 - register (including control registers) information for a specific
node
 - examine arbitrary memory addresses (yeah, right :-)
 
While we're at it, how about a led for each transistor in each chip?
It'd only take what, ~1mil led's for a i860?  Let's see, that's a
1000x1000 pixel grid, and at 400dpi (using Kodak(?)'s new 400dpi lcd),
that's about 2.5x2.5in.

With a 3'x6' panel, we're looking at roughly 15x30 such lcds, giving
us the ability to examine the cpus of 450 nodes of a reasonably large
sigma... Another door's worth would probably give us all the support
chips.  The real problem comes with ram: 1MB of 9-bit parity ram would
take a cell 7"x7" to show each bit... Oh well, maybe just light up a
point based on the parity bit to show bad memory?

This is getting out of hand, I suspect...


--
             J. Eric Townsend --  jet@nas.nasa.gov -- 415.604.4311
               '92 R100R, DoD# 0378, BMWOA, AMA, ACLU, EFF, CPSR
                                "HotHead Rules!"

In article <1992Nov16.212703.4519@nas.nasa.gov>, fineberg@win31.nas.nasa.gov (Samuel A. Fineberg) writes:
|> In article <1e8tm6INN61f@early-bird.think.com>, ephraim@think.com (Ephraim Vishniac) writes:
|> |> In article <1992Nov16.174414.12427@craycos.com> jrbd@craycos.com (James Davies) writes:
|> |> >Actually, even the $250,000/GFLOP number sounds a bit low to me; $2M/GFLOP
|> |> >is closer to the current going rate.
|> |> 
|> |> I don't know what Cray's prices are, but $250K/GFLOP is the ballpark
|> |> figure for the CM-5. 
|> |> 
|> |> -- 
|> |> Ephraim Vishniac    ephraim@think.com   ThinkingCorp@applelink.apple.com
|> |>  Thinking Machines Corporation / 245 First Street / Cambridge, MA 02142
|> |>         One of the flaws in the anarchic bopper society was
|> |>         the ease with which such crazed rumors could spread.
|> Only if you mean peak.  And if you think a CM-5 can deliver $250K/GFLOP on
|> real application code, I have a bridge over the San Fancisco bay you might 
|> be interested in.

A curiosity question. When TMC sells people a CM-5, do they say it will do 1
GFLOP per $250K, or do they actually note that this performance is only
approached by heinously parallel code. (i.e. Are they trying to sell the CM-5
as a general purpose supercomputer?)

Jason W. Solinsky

In article <JET.92Nov17104503@boxer.nas.nasa.gov> jet@boxer.nas.nasa.gov (J. Eric Townsend) writes:
>Blinking lights.  Lots and lots of blinking lights....

Ah, yes, blinking lights.  The first parallel machine I worked on  (C.mmp,
a collection of 16 PDP-11s built at Carnegie-Mellon in the 1970's) had
a pretty fair collection of blinking lights, and looked pretty when
you turned the computer room lights off.  Not only did each PDP-11 have
a front panel with a row of lights, the crosspoint interconnection for the
shared memory had 256 lights, one for each crosspoint.  I have forgotten the
original rationale for putting those lights on the crosspoint, but we
actually diagnosed some memory conflict performance problems by watching the
lights while an application ran.  In addition, the "I'm dead" error routines
from the processors were written to make particular patterns on the front
panel of the PDP-11s.

However, current systems have other methods for getting similar information.
Other than being a designer preference for "flash" vs "cool", what benefits
do lots of blinking lights offer, and what should they represent?
I.e. Do people have suggestions for specific hardware events that might
be connected to blinking lights on a massively parallel shared memory super?

- Patrick McGehearty

In article <JET.92Nov17104503@boxer.nas.nasa.gov>, jet@boxer.nas.nasa.gov (J. Eric Townsend) writes:
|> 
|> 
|> What a silly question.
|> 
|> Blinking lights.  Lots and lots of blinking lights.   well, if one
|> buys a massively parallel supercomputer, one gets blinking lights.

You have the wrong technical term here. Those aren't blinking lights, 
they're the "Front panel visualization tool" (Page 6 Intel Paragon 
Supercomputer glossy : order no 203/6/92/10K/GA)

-- 
James Cownie 
Meiko Limited
650 Aztec West
Bristol BS12 4SD
England

Phone : +44 454 616171
FAX   : +44 454 618188
E-Mail: jim@meiko.co.uk or jim@meiko.com

In article <1992Nov16.174414.12427@craycos.com> jrbd@craycos.com (James Davies) writes:
>In article <1992Nov16.141214.7632@walter.bellcore.com> mo@bellcore.com writes:
>>if you know how to build a 1GFLOP machine for $10,000
>>(much less sell it for that much), then you should do
>>so immediately and become immensely wealthy.
>>
>>i don't know where you got this $10K number, but boy,
>>are you confused.
>
>Actually, even the $250,000/GFLOP number sounds a bit low to me; $2M/GFLOP
>is closer to the current going rate.


MIPs aren't the problem - I/O and memory access is the problem.  Any old 
fool can plug together a bunch of Intel CPU chips and claim that the 
resulting box is faster than a Cray (usually the model isn't specified :-) 
but it takes an expert to get the bus architecture correct.  It's also damn 
expensive.










-- 
Geoff. Lane. | Internet: zzassgl@uts.mcc.ac.uk | Janet: zzassgl@uk.ac.mcc.uts
UTS Sys Admin,  Manchester Computing Centre,  Oxford Rd,  Manchester, M13 9PL
During a week's TV I want to be entertained, bored, delighted, disgusted, made
happy, saddened, enlightened, mystified, else they aren't doing the job right.

In article <1992Nov18.095903.8190@meiko.com>, jim@canopus.meiko.co.uk (James Cownie) writes:
|> In article <JET.92Nov17104503@boxer.nas.nasa.gov>, jet@boxer.nas.nasa.gov (J. Eric Townsend) writes:
|> |> 
|> |> 
|> |> What a silly question.
|> |> 
|> |> Blinking lights.  Lots and lots of blinking lights.   well, if one
|> |> buys a massively parallel supercomputer, one gets blinking lights.
|> 
|> You have the wrong technical term here. Those aren't blinking lights, 
|> they're the "Front panel visualization tool" (Page 6 Intel Paragon 
|> Supercomputer glossy : order no 203/6/92/10K/GA)
|> 
|> -- 
|> James Cownie 
|> Meiko Limited
|> 650 Aztec West
|> Bristol BS12 4SD
|> England
|> 
|> Phone : +44 454 616171
|> FAX   : +44 454 618188
|> E-Mail: jim@meiko.co.uk or jim@meiko.com
Actually, the "Front panel visualization tool" is a tool that lets you see
the blinking lights from a workstation.  I don't know what name they give
the lights themselves.  They probably call them the Blinkium.

Sam

In article <1992Nov17.192804.4410@news.eng.convex.com> patrick@convex.COM (Patrick F. McGehearty) writes:
>In article <JET.92Nov17104503@boxer.nas.nasa.gov> jet@boxer.nas.nasa.gov (J. Eric Townsend) writes:
>>Blinking lights.  Lots and lots of blinking lights....
>
>However, current systems have other methods for getting similar information.
>Other than being a designer preference for "flash" vs "cool", what benefits
>do lots of blinking lights offer, and what should they represent?
>I.e. Do people have suggestions for specific hardware events that might
>be connected to blinking lights on a massively parallel shared memory super?

I'm a great fan of blinking lights. They can be quite valuable for the system
programmers and for knowledgeable application programmers. On the Myrias
system, we have 6 software-controllable LEDs per processor on a PE tile, and
a few others on the communication card to indicate backplane and intercom
cable activity. The current Myrias kernel uses the 6 LEDs for things like:
executing user code, executing system code, executing in BERR mode, TOD tick,
etc. Special patterns are used when something goes worng. Thus, a quick
glance at the lights can show if the user code has crashed the domain.
Hardware problems can show up as procesors whose lights don't follow the
same pattern as others. Performance bottlenecks can show up similarly -
detecting "stragglers" (tasks that run much longer than all others) is quite
easy. Most of these things can also be done without looking at the lights,
but it is often quicker to just take a quick look. I've seen applications
programmers use the lights to get a feel for the kinds of data motions
(in very rough terms) that a program is causing.

In my own case (programming the IOPs), I use the top 2 LEDs as a simple
blinker in my idle loop - how quickly they blink is a nice indicator of
how busy it is. I've used other lights as simple counters for things like
transaction traffic - even a couple of bits can give you some idea of the
rates, etc. by how fast they blink (you have to pick which bits you want
to see of the real counter). A nice combination of hardware-controlled LEDs
and software controlled ones is VERY good to have.

And last, but not least is hardware diagnostics. If the power-up code does
self-tests, it can display its progress on the LEDs, so that it is easy to
get a clue about what is wrong with something. Our board bring-up people also
used this extensively when bringing the boards up originally.

When a remote site complains that something doesn't work, we have often had
them look at various lights and tell us what they are doing.
-- 
Chris Gray	Myrias Computer Technologies Inc.   (403) 463-1337
	cg@myrias.ab.ca          ...!uunet!myrias!cg