Prism workstation is out!

Open discussion on or around SGI; software/hardware related posts should go in the appropriate subforum.
Forum rules
Any posts concerning pirated software or offering to buy/sell/trade commercial software are subject to removal.
User avatar
LaLora
Posts: 258
Joined: Sat Jul 24, 2004 7:54 am

Unread postby LaLora » Thu May 12, 2005 6:56 am

or how many cpus are needed to get the
result.


if all different types of cpus would consume SAME amount of power and if their manufacturing cost would be SAME then you would be right. Unfortunately, you're wrong..

User avatar
quasi
Posts: 92
Joined: Mon Nov 29, 2004 11:42 am
Location: Montréal, Canada

Unread postby quasi » Thu May 12, 2005 2:50 pm

codemonkey wrote:Besides, according to top500.org, an IBM PowerPC based machine called BlueGene is top ranked at a theoretical performance of 91750 GFlops with a peak performance of 70720 GFlops which is almost SGI's / NASA's Columbia AND NEC's Earth Simulator put together!


Note that to get top500, BlueGene needs 3 times the CPU as the NASA Columbia Altixes. Also to use the performance of the IBM, you need stuff that are embarassingly parallel, otherwise your performence drops very quickly, the linpack used by the top500.org is that kind of parallel job.

I was at the supercomputing conference in Pittsburgh when they annonced the top500 list, SGI and NASA people were not happy with the coup IBM pulled. They were quick to point out that the BlueGene was an in LAB system and the NASA/Columbia was a system that was at the client and used.

SGI are still #1 with single image system on several CPUs (512). The next top500 list will be out in June at the ISC in Germany.

TeeTylerToe
Posts: 913
Joined: Mon Sep 13, 2004 11:56 pm

Unread postby TeeTylerToe » Thu May 12, 2005 2:58 pm

isn't the blue gene some sort of experimentation with a supercomputer with a smaller footprint? iirc there was something about a dishwasher configuration. maybe the reason that the blue gene needs more chips, is that they're low power or something.

TeeTylerToe
Posts: 913
Joined: Mon Sep 13, 2004 11:56 pm

Unread postby TeeTylerToe » Thu May 12, 2005 2:58 pm

isn't the blue gene some sort of experimentation with a supercomputer with a smaller footprint? iirc there was something about a dishwasher configuration. maybe the reason that the blue gene needs more chips, is that they're low power or something. hmmmm, 125W 1.6GHz ia64

User avatar
LaLora
Posts: 258
Joined: Sat Jul 24, 2004 7:54 am

Unread postby LaLora » Thu May 12, 2005 3:36 pm

Blue gene is not an experimental project anymore. Blue Gene is now available for sale as a commercial product to everyone who wants it. The full system model name is "IBM eServer Blue Gene" and here is official product webpage with specification pdfs at IBM:

http://www-1.ibm.com/servers/deepcomputing/bluegene.html

It is available as a single rack with (I think) 1024 processors and more racks can be clustered together. The system is exactly the same one as the current fastest supercomputer in the world. Nice cabinet style, by the way, I really like it :) You're very right about that TeeTyler, ..the whole point is to create an extremly massive parallel system with large number of low power consumption cheap cpus to outperform other systems that use virtually "winter heating devices" like Itanium2, vector processors etc. It also has a very small footprint - only 0.9 X 0.9 m (if I remember correctly)..

Had SGI been smart, they could use (or develop) their own cheap low power consumption MIPS RISC processors in large numbers (even for desktops) instead of half dead Itanium2 on which you could fry an egg. Whit all the experience SGI has, one of the choices could be using latest cheap low powered NEC MIPS RISC cpus for embedded systems and pack them together in large numbers - could be possible that such systems would still be cheaper, faster and less "hot" than Itanic..

..however, ..people, don't get me wrong - don't mix hobby and a large scientific department decisions where money, performance and heat generation are essencial (..ok, for me to some extent that goes hand in hand :) ) ..I still want an old small Cray or NEC vector processor system in my home - just for pure fun of owning and operating a rare piece of history :)

jollyroger
Posts: 25
Joined: Fri Oct 01, 2004 4:00 pm

Unread postby jollyroger » Fri May 13, 2005 1:15 pm

As LaLora pointed out performance is ultimately essential, and so is power consumption and heat dissipation.
The smaller footprint of BlueGene makes it capable of using more computing processors in a unit of space.
On top of that the kind of computation the scientific community needs uses pretty much always massively parallel algorithms, which are suitable for a cluster of many low power processors.
In the end it is just much, much more powerful, when it is completed BlueGene-L is going to be several times faster than Columbia I am afraid...

User avatar
foetz
Moderator
Moderator
Posts: 6593
Joined: Mon Apr 14, 2003 4:34 am
Contact:

Unread postby foetz » Fri May 13, 2005 3:56 pm

LaLora wrote:
or how many cpus are needed to get the
result.


if all different types of cpus would consume SAME amount of power and if their manufacturing cost would be SAME then you would be right. Unfortunately, you're wrong..


sure, this would be the economical point but it's about performance, right?
so if the altix can get somewhere near the gene with 1/3 cpu count it's the better machine.

User avatar
foetz
Moderator
Moderator
Posts: 6593
Joined: Mon Apr 14, 2003 4:34 am
Contact:

Unread postby foetz » Fri May 13, 2005 3:59 pm

jollyroger wrote:when it is completed BlueGene-L is going to be several times faster than Columbia I am afraid...


ah this is just future stuff. let's stay with current 'real' things.
who knows what other companies will have in the future...

User avatar
R-ten-K
Posts: 1889
Joined: Mon Nov 15, 2004 10:36 pm
Location: Nor Cal

Unread postby R-ten-K » Sun May 15, 2005 9:21 pm

foetz wrote:so if the altix can get somewhere near the gene with 1/3 cpu count it's the better machine.


Nope....

unixmuseum
Posts: 2783
Joined: Mon Apr 19, 2004 4:25 pm
Location: Los Angeles, CA

Unread postby unixmuseum » Sun May 15, 2005 11:16 pm

Dubhthach wrote:Given the heat output of Itaniums that shouldn't be too surprising ;) no doubt it was also cheaper for them to develop as they didn't have to do too much in way of chipset design. Still the next version of Itanium is dual core and is pin compatible with those, so within 9months or so ye be able to upgrade your 2CPU Prism workstation 4CPU (2 x dual core)

Yes, the Itanium is going to get a dual core update next year, the prism is also going to be quad CPU in the near future, maybe "quad cpu" means using 2 dual core Itanium... There is apparently a lot of work being done on the next gen Itanium, at least according to HP & Intel... I got a mixed message from Intel, as if they want Itanium to be both cutting edge and dead at the same time... Nonetheless, HPC benchmark results still show Itanium-based hardware scaling up much better than x86-64...

User avatar
foetz
Moderator
Moderator
Posts: 6593
Joined: Mon Apr 14, 2003 4:34 am
Contact:

Unread postby foetz » Wed May 18, 2005 11:02 am

R-ten-K wrote:
foetz wrote:so if the altix can get somewhere near the gene with 1/3 cpu count it's the better machine.


Nope....


why?

TeeTylerToe
Posts: 913
Joined: Mon Sep 13, 2004 11:56 pm

Unread postby TeeTylerToe » Wed May 18, 2005 1:20 pm

I may have this wrong, but blue gene performance == altix performance

but the ratio for performance per processor favors the altis... so why isn't the altix better.

1. TCO. the altix uses much much more electricity, and it ain't free

2. footprint (physical) I suppose this is included in tco, but it's big, and so it's more expensive to house

3. thermal footpring, another tco thing I suppose, but you've also got to cool that altix.

so assuming that performance is equal, the altix is less attractive.

User avatar
foetz
Moderator
Moderator
Posts: 6593
Joined: Mon Apr 14, 2003 4:34 am
Contact:

Unread postby foetz » Wed May 18, 2005 1:24 pm

TeeTylerToe wrote:I may have this wrong, but blue gene performance == altix performance

but the ratio for performance per processor favors the altis... so why isn't the altix better.

1. TCO. the altix uses much much more electricity, and it ain't free

2. footprint (physical) I suppose this is included in tco, but it's big, and so it's more expensive to house

3. thermal footpring, another tco thing I suppose, but you've also got to cool that altix.

so assuming that performance is equal, the altix is less attractive.


maybe but we've had the economical issue already.
i just said the altix is the better machine. i did NOT say it's cheaper or smaller etc.

hope it's clear now.

User avatar
foetz
Moderator
Moderator
Posts: 6593
Joined: Mon Apr 14, 2003 4:34 am
Contact:

Unread postby foetz » Wed May 18, 2005 1:27 pm

additionally one can look at other features like bandwidth etc. (gene's interconnect is a joke).
also consider the max. cpu number per image!

User avatar
R-ten-K
Posts: 1889
Joined: Mon Nov 15, 2004 10:36 pm
Location: Nor Cal

Unread postby R-ten-K » Wed May 18, 2005 3:42 pm

foetz wrote:
R-ten-K wrote:
foetz wrote:so if the altix can get somewhere near the gene with 1/3 cpu count it's the better machine.


Nope....


why?


Because Columbia will not be anywhere near what the BlueGene/L at LLNL offers flop wise, so that was a moot point.


Return to “SGI: Discussion”

Who is online

Users browsing this forum: No registered users and 1 guest