Cutting and pasting and shifting from my comments in the thread in GH:
The article is interesting because it starts off with a spurious premise ("Intel's new Itanium 2 super-chip is said to be ideal for use on search engines like Google.") and builds on this to imply that Itanium 2 is a failure - or will be in the future. I like the "is said to be" - which automatically leads me to ask "is said by whom?". Google not an ideal candidate for an Itanium 2. They are an ideal candidate for a blade processor system. If they aren't into blades already, that's where I would automatically think that they should be.
I think that Intel is pleased with the success of Itanium 2 sales so far. But the article could substitute any of the super chips out there: Compaq's Alpha, IBM's Power 4, Intel's Itanium, and others and say that their power density is too high. You could substitute "Intel" with "IBM and "Itanium 2" with "Power 4" throughout the article and come up with the same type of article. Which could lead to the false conclusion that there are no uses for super powerful server/workstation processors because distributed computing solutions are more economical. There are plenty of solutions where a distributed computing approach simply can't be implemented. And when you can't distribute the problem, and you want to get it done quickly, then a high-end processor is the solution that you want no matter what the power density is.
Commenting more on why I don't think Google is a good candidate for an Itanium 2 solution: Google doesn't talk much about exactly how they work, but the general gist of is that requests are sent to a processor for lookup in a relatively static datebase. The database is "static" because a search does not immediately change the results of other searches... it does happen, but not immediately. What they want for these types of things are a whole lot of cheap, moderately fast, processors - like a Beowolf cluster. Compare this situation to a slightly different one where you are managing a credit card database. In this case, the results of one transaction affect the results of any future transactions. In this case a Beowolf-like cluster of distributed machines will not scale well and what is needed is faster processors working in some form of large server box - like a 32-way machine. In these cases, speed is everything. In Google's case, speed is not everything. Cost probably is.
Facebook
Twitter