wbmw, you seem to fundamentally misunderstand redundancy -
They will care, though, if important information is returned. If the information financial, medical, or personal, then an error would be disastrous. How would you like your $50,000.00 bank account reduced to $10,000.00? How would you like the doctor to give you the wrong pills and make you even more sick?
The concept of redundancy is to give accurate results even when components fail. You probably know this from ECC DRAM - if a bit is bad, then it is reconstructed from the two redundant bits in the 10-bit byte.
Similar mechanisms exist on the CPU or system levels. One traditional method is voting - have three systems working on the same problem, and if one gives an answer different than the other two then the two that agree are considered correct. With the proper selection of different applications running across the total array of machines you make sure that no two machines is running the same suite of programs, giving a measure of insurance against a application error due to a bug interacting with the system environment.
A simpler method is to use two systems and, if they disagree on the answer, throw away both & give the problem to a new set of two systems.
No matter what system of redundancy is used, transactions are not completed to the customer until it is recorded in the computer. You never give an ATM permission to release funds until the customer's account has been debited, for instance.
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. To me, the Itanium approach is not a practical solution to this problem. Itanium is a number cruncher that is good for certain applications, but robustness won't be one until the CPUs are way, way cheaper so that they can be easily made redundant.
Market Data 
Markets 
AI
