Mainframes are overpriced and inefficient, but they are the only option for a F500 without the in-house talent to build any kind of distributed, fault tolerant system.

But people have been predicting the death of big iron for several decades now, yet they live on.

I don't doubt mainframes might be overpriced, but I also suspect the reason they persist is they have yet to come up with a cheaper option, offering the same performance figures.

Also, there is a remark that in major parts of the financial industry, running an IBM mainframe is nearly a necessary condition for compliance.

I'm don't doubt that is a major factor. Add to that the major risk that what every new system you move to might actually fail to work or end up costing more.

But isn't that what Facebook, Google, and Amazon are doing? Using massively distributed commodity x86 hardware to eat away at incumbent businesses that would outsource their IT services to mainframes? Last I read, Google is about to go into auto insurance, and all three companies I listed do payment processing.

If software is eating the world, SV behemoths are eating business verticals.

[1]

[1] http://www.share.org/p/bl/et/blogid=2&blogaid=234

Second, since I assume, most of the programmers here don't have access to mainframes, it is hard to testify those numbers. Also, since it is a benchmark it will be useful to revel what exactly the task they are using here, otherwise I would simply throw this claim into my 'pure PR mess, don't take it seriously' bin :)

One use of CICS was for heads down medical claims processing, across all four US time zones. The site our team from IBM Research visited wanted high reliability: If the site was down for, say, an hour, then the data entry staff would have to be called back on a Saturday, for at least half a day, at a higher rate per hour. One such outage in a year, and the CIO could lose his bonus. Two and he could lose his job. The site was very uptight. Getting into the glass house was not easy; might have been easier to get into the White House Oval Office.

At one time to make CICS more secure, there was some interest in having processor hardware support for address sub-spaces. Another idea was cross memory where a program could call and execute, say, a function in another address space. There were also data spaces, that is, address spaces with just data and no code but that could be accessed by other address spaces with code.

Net, the IBM mainframes are not really simple things. Cloning one would not be easy, and at IBM's next version of hard/software, the clone could be unable to run the newer software and suddenly be a boat anchor.

The mainframes have their own CISC type architecture and typically have massive amounts of cache and a high clock speed (5Ghz and above)

Their instructions sets are also different in that its not strictly of the von-Neumann variety. Mainframes can do things like memory copies directly in memory without requiring copying via registers. This kind-of attacks the von-Neumann bottleneck directly and is good for batch and high volume transaction processing.

The software for mainframes is also typically fused into their kernels, things like CICS and DB2 are not "user" programs, they're part of the OS so I/O is handled much better and things don't "block" as much.