Sgolds, Re: Bottom line - it is less expensive and more robust to insure accuracy by building redundancy with cheaper component systems than to architect a single, expensive system to do the same.
Not true. Your examples required 2-3x the hardware to do the same as an integrated, built in RAS feature. If scaling out were really as easy as it seems, then builders would be making server farms out of VIA C3 chips and cover whole floors of buildings with all the hardware.
However, no one wants 3x more systems than they need. It raises upkeep costs, takes up more space, and increases the chance that some systems go down. You don't expect them to just pull systems from their racks every time one goes down, do you? Such a system would be very poorly managed.
Consider the IBM commercials where one server replaces the functionality of an entire room of servers. Some companies like having their data local - at a single point in failure. Not traveling across multiple systems in a redundancy failover mode. What happens if you have to service a machine and take it off the node? Then you leave yourself exposed. High RAS machines allow uptime while they are serviced.
Also, in high bandwidth situations, scale-out machines are slower than scale-up machines. Nothing beats having all the data traveling on high speed internal interconnects, rather than on cluster connections. Itanium with multiple cores will have several times the transaction processing as x86. If the future Tanglewood processor is built OOOE (quite possible with the Alpha designers), it could run circles around x86 in scalar apps, as well.
The past couple years have shown me that so far, the Itanium critics have been dead wrong about this architecture. They thought it would fail miserably as an underperformer, but now it's the fastest scalar processor on the planet. They thought that Dell and IBM would decommit from Itanium, and now both are on board. They thought software would never come, and now there are hundreds of apps on all the major operating systems. They thought that 32-bit performance would be years behind x86, but the upcoming IA-32 EL will let it catch up to decent speeds. All the doom and gloom has proven to be groundless, and I see no major blockers for Itanium ramping to be the industry's next major architecture. It will take several years, but the roadmap is much stronger than x86, and Intel is investing big bucks into the software and infrastructure.
Opteron's paradigm is just not convincing enough, IMO.
Market Data 
Markets 
AI
