🚀 go-pugleaf

In article <1992Aug11.214207.16247@tera.com> bob@tera.com (Bob Alverson) writes:
>Having just read Gordon Bell's article on supercomputers in the latest
>CACM, I have to ask: Is Bell financially associated with KSR?

Bell and Henry Burkhardt III (founder, and until recently, chairman of KSR)
were both involved with Encore. I believe they were both founders.  During
the period I worked at KSR, Bell showed up several times. On the other
hand, I believe Bell also showed up at Thinking Machines from time to time.

I don't know about any financial association Bell has with KSR or TMI.

There was an article in a magazine called (I believe) "Upside" in the last
year. The magazine appeared to be oriented towards technology-following
venture capitalists. My recollection of the opinions from that article
(glad to hear any others):
 - the original Connection Machine architecture was neat, using many many
   simple processors
 - the KSR architecture is neat, using a smaller number of more complex
   processors
 - the CM5 is a betrayal of the original CM architecture
 - insults were offered by Bell, Burkhardt, and Hillis
 - some other heavy (Grosch?) was quoted; I believe also anti-CM5
 - there is some kind of "bet" (involving Bell) about which machine type
   will triumph

I found the opinions expressed in the article to be amazingly "hard"; I
didn't see that they were justified. (However, I'm no visionary, and there
aren't any giants nearby who will let me stand on their shoulders.)
Certain people like to make wild long-shot predictions and
characterizations: if they are right, then the people point back to the
prediction, and are dubbed prescient; if they are wrong, well... who's
counting how often they are wrong?

There have also been several articles about KSR in the Boston Globe
business section where somewhat similar opinions were voiced.

According to the quarterly report I just received, KSR has delivered the
second level interconnection and has run a system of 128 CPUs. I believe
Cornell will have two 32-processor groups connected by a second-level
interconnect.

(I obviously have some financial interest in KSR, but nowhere close to
the America-3 level...).

-----
E. N. Kittlitz                                          kittlitz@world.std.com
contracting at OSF, not representing their positions.   kittlitz@osf.org

What I found amusing in Bell's article was the claims that KSR had proven
that its architecture was scalable to 1088 nodes.  At a talk I heard they
had never tried constructing a machine with more than 128 nodes, and have
little experinece with machines larger than 64 nodes.  Thats a long way
from 1088.  It reminds me of the claims TMC has made about the CM5's 
scalability to 16K nodes.

Sam

In article <1992Aug12.171629.2858@nas.nasa.gov> fineberg@nas.nasa.gov writes:
>What I found amusing in Bell's article was the claims that KSR had proven
>that its architecture was scalable to 1088 nodes.  At a talk I heard they
>had never tried constructing a machine with more than 128 nodes, and have
>little experinece with machines larger than 64 nodes.  Thats a long way
>from 1088.  It reminds me of the claims TMC has made about the CM5's
>scalability to 16K nodes.

In general, the hype in the MPP field is thick enough to cut with a
knife.  Unfortunately, many large organizations are getting swept up by it and
are basing purchasing decisions on it.  The result could well be that time on
supercomputers capable of addressing the needs of the bulk of current users
may be hard to come by in a few years.  Should be good news for the
workstation manufacturers and people who are selling workstation networking
software (along those lines, there was an interesting article in Monday or
Tuesday's Wall Street Journal about people abondoning supercomputers in favor
of networked workstations). 

John





-- 
Dr. John K. Prentice                      
Partner, Quetzal Computational Associates    
3200 Carlisle N.E., Albuquerque, NM   87110-1664; 505-889-4543
john@unmfys.unm.edu -or- jkprent@cs.sandia.gov -or- prentice@rufous.cs.unm.edu

In article <1992Aug11.214207.16247@tera.com>, bob@tera.com (Bob Alverson) writes:
:
|> Also, there appears to be an error in Table 1b of the article.  It lists
|> the KSR 1 as achieving 513 Gflops on Linpack 1000, with 32 processors.
|> Maybe it should be 513 Mflops--does anyone know for sure?
:

The table is clearly mistaken, because in the previous table, it is clear
that each processor is capable of 2 floating point operations per clock, 
and there are a maximum of 1088 processors at 20 MHz.  32 processors would
give a peak of 1.28 GFLOPS.  So, the listed 513 GFLOPS is clearly an
error.  513 MFLOPS would be consistent, although I don't know if it is
correct.


-- 
  Hugh LaMaster, M/S 233-9,     UUCP:      ames!lamaster
  NASA Ames Research Center     Internet:  lamaster@ames.arc.nasa.gov
  Moffett Field, CA 94035-1000  Or:        lamaster@george.arc.nasa.gov 
  Phone:  415/604-1056                     #include <usenet/std_disclaimer.h> 

In article <1992Aug12.171629.2858@nas.nasa.gov> fineberg@nas.nasa.gov writes:
>What I found amusing in Bell's article was the claims that KSR had proven
>that its architecture was scalable to 1088 nodes.  At a talk I heard they
>had never tried constructing a machine with more than 128 nodes, and have
>little experinece with machines larger than 64 nodes.  Thats a long way
>from 1088.  It reminds me of the claims TMC has made about the CM5's 
>scalability to 16K nodes.

A similar early claim was made for the CMU Cm* as being capable of having
almost 16K processors (50 were made, just a matter of how much you wanted
to spend).  You can tell how many bits were used for processor
addressing using this number (right?).  That's "Design" and "Engineering."
"Proving" you can build it is "Manufacturing." 8^)

Steve Lundstrom over at Stanford believes that unless you build to the maximum
designed size then sell off the pieces, you won't discover all the problems
(i.e. physical world) inherent in a design.  Other "weird" [typically not
powers of 2] numbers have their reasons in the world (the PEPE had a weird
number, the Delta has a non-2^n number); you can only ask why and hear their
explanation.

--eugene miya, NASA Ames Research Center, eugene@orville.nas.nasa.gov
  Resident Cynic, Rock of Ages Home for Retired Hackers
  {uunet,mailrus,other gateways}!ames!eugene
Second Favorite email message: Returned mail: Cannot send message for 3 days
A Ref: Mathematics and Plausible Reasoning, vol. 1, G. Polya

In article <!14m#9l@lynx.unm.edu> john@aquarius.unm.edu (John Prentice) writes:
>In general, the hype in the MPP field is thick enough to cut with a
>knife.  Unfortunately, many large organizations are getting swept up by it and
>are basing purchasing decisions on it.  The result could well be that time on
>supercomputers capable of addressing the needs of the bulk of current users
>may be hard to come by in a few years.  Should be good news for the
>workstation manufacturers and people who are selling workstation networking
>software (along those lines, there was an interesting article in Monday or
>Tuesday's Wall Street Journal about people abondoning supercomputers in favor
>of networked workstations). 

Two years ago at the yearly Hackers Conference, there was some discussion
brought up about distributed programming after Hillis spoke about
genetic algorithms.  The general consensus appeared to be that programmers
did not care for centrally located resources whose time had to be scheduled or
where resource conflicts might take place.  While this view was a little
naive, I think it shows that people want to control their own resources.
[It's only naive because most of the people expressing the opinion were
not thinking about system balance and were not people aware of the state
of parallelism software.]  I believe it will go on pushed by things like PVM.

--eugene miya, NASA Ames Research Center, eugene@orville.nas.nasa.gov
  Resident Cynic, Rock of Ages Home for Retired Hackers
  {uunet,mailrus,other gateways}!ames!eugene
Second Favorite email message: Returned mail: Cannot send message for 3 days
A Ref: Mathematics and Plausible Reasoning, vol. 1, G. Polya

The KSR1, delivered to the Cornell Theory Center late in '91 and put
through Acceptance in May '92, did indeed achieve 513 MFLOPS for a
1000x1000 Linpack, and was informally reported to have achieved
~517 MFLOPS on an in-house machine (in Waltham) prior to the actual
Acceptance Test performed here on ours.

As previously posted, we are at this time awaiting final tuning of our
64-node machine (2 rings of 32 processors with a level-1 interconnect,
in KSR-speak "a Search-Group 1 comprised of two Search-Group 0's")
before beginning the Acceptance Test for that system.

doug
--
#  ____    |Internet:   elias@.tc.cornell.edu
#dr _|_)oug|USmail:     Sci.Comp.Support/Cornell Theory Center
#  (_|     |            737 TheoryCtrBldg/C.U./Ithaca/N.Y./14853-3801
#  (_|__   |MaBelle:    607-254-8686   Fax: 607-254-8888

In article <ELIAS.92Aug20095752@fitz.TC.Cornell.EDU> elias@fitz.TC.Cornell.EDU (Doug Elias) writes:
>The KSR1, delivered to the Cornell Theory Center late in '91 and put
>through Acceptance in May '92, did indeed achieve 513 MFLOPS for a
>1000x1000 Linpack, and was informally reported to have achieved
>~517 MFLOPS on an in-house machine (in Waltham) prior to the actual
>Acceptance Test performed here on ours.
>

Yesterday I heard some very similar performance numbers for a CFD code run on
about the same size KSR system.  

John

-- 
Dr. John K. Prentice                      
Partner, Quetzal Computational Associates    
3200 Carlisle N.E., Albuquerque, NM   87110-1664; 505-889-4543
john@unmfys.unm.edu -or- jkprent@cs.sandia.gov -or- prentice@rufous.cs.unm.edu