Post a Comment
I wonder how long it is going to take Intel to completely fold Itanic... I give it another 3 years before Intel burns through another $2Bn in wasted investment on Itanic and completely pulls the plug. It is going to be interesting what happens to SGI then. If I were making decisions at SGI, I would be seriously considering Opteron now.
RE: Itanic can kill SGI
>Everyone is doing their data on grids and stuff now
Yeah like the terragrid folks;
http://www.gridtoday.com/05/0117/104473.html
no need for that big ol' iron any more 
I think you may be somewhat wrong on this if this is your 'general view'. In a large number of cases, the intra-node communication becomes the bottleneck - even on clusters with Myrinet, Dolphin SCI or other hi speed interconnects. In those settings, grids are not really going to help you much.
That said, there are plenty of 'batch processing' tasks that runs very, very well on grids, e.g. finance/monte carlo simulations are great for that.
They're offordable right now; in amounts of 1000, they're priced only slightly higher than the Xeon; the problem is, however, you can't buy the CPU's off any distributor, and no third party motherboard and component manufacturers are providing solutions to small white box computer businesses, which still make up over 40% of the market place - in some countries, higher.
Supercomputing with regular off the shelf commodity chips is real wasteful.
We need to use streaming processors for alot of projects. They are more efficient.
http://merrimac.stanford.edu/
I do like itanium though.
Anonymous,
Would you care to elaborate ? "Supercomputing with regular off the shelf commodity chips is real wasteful" does not provide much insight as to which industries and purposes you are talking about.
I certainly agree about the need for streaming processors, but by and far, they (possibly with the exception of ClearSpeed) are not mature enough for most industry specific purposes (in large scale settings, not thinking of 'desktop tinkering')
Too expensive and running on dead end intel cpus. They feel the squeeze from the desktop sector so they must lower prices on their hw and they make it then look like they're doing us a favor by dropping prices on their crap hw. You can probably build a rack server from desktop computers and use some linux sw to run it at fraction of the cost and faster.
For that kind of hardware, the price really is a bargain. I believe that Sun is offering competitive hardware with slightly better prices, but Iīm not sure. And to anyone who thinks that one could build a rack of off-shelves desktop parts and just throw Linux at it (letīs pretend that those 1U small factor mobos and cases are commodities), I have a quick question. Have you EVER stepped on a Data Center? :-)
I, being an avid Linux user, agree that Linux plays a big part on the performance of these SGI offerings, but puhleese...
However, like some posters already pointed out, Iīd keep both eyes on Intelīs roadmap for Itanium since they arenīt exactly best sellers and everybody and their mothers knows that Intel already spent a lot of money in this. Even HP, Intelīs partner on R&D for the Itanium, already jumped off the wagon.
It would be wise if SGI started to evaluate AMD offerings for its Altix product lines just as a fail safe (or even consider completely replace Intelīs processors with AMD ones since Intel has been playing catch-up to AMD for quite some time).
This is what SGI is really talking about. At the bankruptcy sale, pennies on the dollar, so that will be $70 per 1U supercomputer.
$70 is not a bad deal , all things considered. A decent machine with good FP performance and a museum piece as well. At long last, SGI figured out how to offer hardware at a good price. Hurrah.
Itanium gets a LOT of flak, some of which is deserved, some of which is not.
As it stands, Altix is INCREDIBLY scalable - it beats hybrid-cluster systems such as the Cray XD1 or commodity x86/Myrinet clusters into dirt for almost linear scaling with cpu/job size.
However, the two big issues with itanium are compilers and proce. When we had an Itanium2 system to test, the version of ICC at the time realy struggled in some cases - sometimes the output codes were blazing fast, sometimes they were rubbish compared to a P3 of the same clock speed. When you consider what you can buy for your money, Itanium is simply too risky for uncertain performance characteristics - Ģ20 buys you a 16-core opteron machine, which is more "conventional" to work with, and easier to get reasonablew performance from.
Is ICC better now? probably. But that's no guarantee, and a smart cluster admin, is going to think twice about sinking a lot of money into Altix.
http://www.primidi.com/2003/12/01.html
a little on streaming processor for supercomputers.



