Author: Brian Richardson
Date: 07:07:39 01/11/02
Go up one level in this thread
On January 10, 2002 at 22:05:54, Vincent Diepeveen wrote: >On January 10, 2002 at 10:33:37, Brian Richardson wrote: > >>On January 10, 2002 at 07:48:26, Ed Schröder wrote: >> >>>http://www.cnn.com/2001/TECH/industry/12/24/ibm.bluestorm.idg/index.html >>> >>>For the record, Blue Storm will contain 1.5PB (petabytes) of data by 2004, equal >>>to 75,000 20GB PCs. And it will crunch 20 trillion calculations per second, >>>meaning that if this reporter used a hand-held calculator to race Blue Storm to >>>20 trillion calculations, I would be running behind by 17 million years, >>>according to IBM. >>> >>>Also, chess players be warned, Blue Storm at its apex, will be 1,700 times >>>faster than Deep Blue, the super computer that beat chess champion Garry >>>Kasparov, IBM officials said. >> >>There are many ways to measure "speed". While I have not looked at it in detail >>yet, I doubt that it will be 1,700 faster than DB for chess. However, it is >>interesting since the basic node is the p690, which can have up to 32 way SMP >>(1.3GHz). There are many configuration options utilizing IBM's Power4 CPUs, >>some with shared L2 cache (Power4 enables 2way on chip SMP), some with dedicated >>L2 and only one CPU per chip. They are packaged in MCM modules (much like >>mainframes), and each module has L3 and memory speed about 13GB/sec (to up to >>256GB of memory). Thus each node is pretty much shared memory. Then you have >>to look at clustering, and it is unlear how many nodes will be supported. > >running on one node is cool thing, but more than one node then also >gigabytes of data a second is not enough for communication speed. > >you can get of course a great speedup on a single node, but that's it >simply. then speedup stops. The latency for communication between >the nodes is going to be just *too* slow. I think we can assume that the interconnect between the large SMP nodes will be _faster_ than the DB SP2 switch... > >32 x 1.3 Ghz is real interesting. Then we of course do not know >what such a processor is. Usually it is slower than PC processors are >for chess. > >Is there any crafty benchmark about these processors? > >If not it's probably a processor in IBMs dreams :) > IBM is not dreaming. The SPECint2000 is 790 overall, but only 674 for the Crafty test. This is about the same as 2GHz Intel. The AMD 1900+ is 940. Note that this is only using 1 CPU of the 2 on the chip, so cache is not shared. Some vendors (Sun) have made an issue of it. So, perhaps we should assume 16 CPUs, not 32 per node. >Any URL to technical documentation about the processor, if i would >guess now i would guess it to be P2 speed at 1.3Ghz and not K7 speed... > Take a look at http://www-1.ibm.com/servers/eserver/pseries/hardware/whitepapers/power4.html >Nevertheless 32 x 1.3Ghz with shared memory is going to give a decent >speedup if all the 32 processors have SHARED memory with all the >processors. > >Estimating the price of a 32 shared memory 1.3Ghz node is pretty hard >though, so is delivery time. IBM is delivering now, with full volume production by March'02. 4 CPU module is $140K, and $160K per 32GB memory (up to 256GB total), plus frame and misc components, of course. A dual CPU per chip 32 way system with 128GB is rougly $2M. > >At the time it gets delivered we already have a dual 5 Ghz or something >available at the stores? > >32 x 1.3Ghz sounds a lot less impressive then :) > > > >>> >>>============ >>> >>>I am sold, have ordered the thing with my VISA card. >>> >>>Anxious waiting for delivery, start porting Rebel, increase the maximum depth to >>>256 ply. The end of computer chess. >>> >>>Ed
This page took 0.01 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.