Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Cray

Author: Robert Hyatt

Date: 12:57:36 07/09/03

Go up one level in this thread


On July 09, 2003 at 00:09:03, Vincent Diepeveen wrote:

>On July 08, 2003 at 19:37:48, Jeremiah Penery wrote:
>
>>On July 08, 2003 at 08:37:49, Vincent Diepeveen wrote:
>>
>>>On July 08, 2003 at 00:33:09, Jeremiah Penery wrote:
>>>
>>>>NEC Earth Simulator has 5120 NEC SX-7(?) vector processors.  Total cost was less
>>>>than $400m.
>>>
>>>around $680M it cost.
>>
>>Provide a reference for that $680m number, and I might believe you.  I don't
>>accept random numbers without reference.
>>
>>Less than $400m is quoted at these sites:
>>http://www.mindfully.org/Technology/Supercomputer-Japanese23jul02.htm
>>http://www.siliconvalley.com/mld/siliconvalley/news/editorial/3709294.htm
>>http://www.time.com/time/2002/inventions/rob_earth.html
>>http://www-zeuthen.desy.de/~schoene/unter_texte/texte/sc2002/tsld004.htm
>>http://www.iht.com/articles/98820.html
>>http://cospa.phys.ntu.edu.tw/aapps/v12n2/v12-2n1.pdf
>>etc., etc.
>>
>>The highest price I've seen is around $500m, nowhere near your number.
>>
>>>>Here is a blurb about the chip, from the webpage:
>>>>
>>>>"Each AP consists of a 4-way super-scalar unit (SU), a vector unit (VU), and
>>>>main memory access control unit on a single LSI chip. The AP operates at a clock
>>>>frequency of 500MHz with some circuits operating at 1GHz. Each SU is a
>>>>super-scalar processor with 64KB instruction caches, 64KB data caches, and 128
>>>>general-purpose scalar registers. Branch prediction, data prefetching and
>>>>out-of-order instruction execution are all employed. Each VU has 72 vector
>>>>registers, each of which can has 256 vector elements, along with 8 sets of six
>>>>different types of vector pipelines: addition/shifting, multiplication,
>>>>division, logical operations, masking, and load/store. The same type of vector
>>>>pipelines works together by a single vector instruction and pipelines of
>>>>different types can operate concurrently."
>>>>
>>>>Each chip consumes only about 140W, rather than Vincent's assertion of 150KW.
>>>
>>>the 125KW is for Cray 'processors' not fujitsu processors that are in the NEC
>>>machine.
>>>
>>>Ask bob i remember he quoted 500 kilowatt for a 4 processor Cray. So i divided
>>>that by 4.
>>
>>That 500KW was probably for the entire machine.  Each processor probably
>
>Yes a 4 processor Cray.
>
>Just for your own understanding of what a cray is. it is NOT a processor.
>It is a big block of electronics put together. So no wonder it eats quite a bit
>more than the average cpu.
>
>That's why i say that those power consuming Crays are history. They are just too
>expensive in power imho. If we then compare that they run at 1Ghz and can do
>like 29 instructions with 256 KB cache, then it is trivial why those matrix
>wonders no longer are a wonder.
>
>Opterons, Itaniums. You might call them expensive in power. It is trivial that
>they are very fast compared to a Cray when you compare the power consumption.
>
>A special water central was typically used to cool those vector Crays. Bob can
>tell more about that. He has had one there at his university.
>
>>consumes a very small amount of that.  The Earth Simulator uses some 7MW of
>>power in total, though only about 10% comes from the processors.
>
>The typical supercomputer has a fast i/o and big routers. Those always eat
>trivially more power than the cpu's.
>
>7 MW nevertheless is hell of a lot.
>
>From chess viewpoint the only interesting thing is what is the one way pingpong
>latency time of the Earth Simulator at the big partitions which work with either
>MPI or openmp. Doesn't matter what of course. Of course not from processors near
>each other but with some routers in between them ;)
>
>Another major difference with Cray machines (using cray processor blocks) is
>typically not using too many processors, because all processors are cross
>connected with very fast connections. No clever routing system at all. Brute
>force.\

Pure cross-bar, the best routing there is.


>
>If you want to make a supercomputer which is having big partitions of cpu's you
>need somewhere a compression point where n cpu's compress to a single bottleneck
>and then with some kind of router or special designed NUMA flex (that's the very
>fast SGI thing where they connect boxes of 64 processors to each other with).
>
>Cray never accepted such bottlenecks. It was just raw vector power. If you
>consider *when* those machines were constructed it was really a genius thing.
>
>It's only now that cpu's are so very well designed and high clocked with many
>instructions a clock that those vector blocks can be replaced safely.
>
>Note i bet they still get used because most scientist know shit from programming
>and you can't blame them.

Sorry, but a Cray will blow the doors off of _any_ microcomputer you care to
march up.  It can sustain a ridiculous number of operations per cycle.  IE it
is _easy_ on a single CPU to add two 64 bit floats, multiply the sum by
another 64 bit float, add that to another 64 bit float.  And I can do all of
that, two results per clock cycle, _forever_.

You have to understand vector processing first, to understand the power of a
Cray.  Until you grasp that, you are talking nonsense.

>
>Today i spoke with someone who is running jobs a lot. What he calls a small job
>is a calculatoin at 24 processors that runs for 20 hours just doing floating
>point calculations.
>
>His software runs already for like 20 years or so at supercomputers.
>
>There is however some major differences with today and back then, that's why we
>spoke. I had promised him to help him speedup.
>
>What he is doing is that a processor has huge 3 dimensional arrays where he gets
>data from.
>
>Those are however allocated at the first thread that starts.
>
>So imagine that 1 poor thread is eating up all that bandwidth of the machine and
>that each cache line to get there takes like 5 microseconds or so to arrive.
>
>Then he can do 16 calculations (cache line length: 128 bytes divided by double
>size = 8 bytes). That's sick expensive.
>
>His software can be speeded up *quite* a lot.
>
>Trivially he ran also in the past at Crays with this software (nowadays it's in
>C, previously it was in fortran).
>
>They just do not know the bottlenecks of todays supercomputers.
>
>That's why the Cray for them was a great thing and always they will remember it
>for that.
>
>Because if you got a processor or 16 with shared memory and for every processor
>a lookup in that memory is equally fast, then it is trivial that this program,
>which definitely is a good example of how many programs still are, can be
>speeded up like 20 times easily at this SGI supercomputer.
>
>Yet the brute force of the Cray doesn't distinguish. So the Cray computer is
>even greater if you realize the average guy who has to do calculations on those
>machine.
>
>Up till recently more than 50% of the total system time goes to researchers who
>are doing physics (if that's the right english word). Calculation of models and
>oil simulations and bunches of known algorithms and unknown new ones that get
>tried with major matrixes.

False.  They are used to design other microprocessors.  Apple owns several.
They are used for weather forecasting.  Simulations.  _anything_ that requires
incredibly high operations per second on large data arrays.  NUMA just doesn't
cut it for many such applications, and message-passing is worse.

_that_ is the "world of the Crays" and they are untouched there.

>
>In this case it was field calculations. Most of the researchers are already so
>happy that they can run in parallel on a machine that we'll forgive them that
>they do some stuff wrong.
>
>In all cases they draw the conclusion that the cpu is eating up the system time,
>because even if your program is 99% busy with calling cache lines from some
>remote node, the 'top' is showing that processes are busy 99.xx% of the system
>time.
>
>let's quote Seymour Cray:
>  "If you were plowing a field, which would you rather use?
>   Two strong oxen or 1024 chickens?"
>
>It's trivial that only the best programmers on the planet can go for that 1024
>chickens.
>

And for a good programmer, those two oxen are going to win the race.

>
>
>>>Trivially Cray machines using the opterons will be consuming less than that.
>>>Note that the cpu costs is nothing compared to what the routers etc eat.
>>
>>Of course.



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.