IBM has made it perfectly clear - buy their x86 based servers.

DB2 is thread sensitive and you can buy 2.2 x86 threads for the same price as each Power 6 thread you purchase, essentialy geting over 1.5 times as much real world work done.

The faster things go the worse the scaling, witness the Pentium 4. Both the Power 6 and the Pentium 4 are single issue cores, starting work on one instruction each clock cycle or a 4.7 Gig raw instructions a second on Power 6.

An Opteron is a three instruction per cycle machine, but a 2.8 GHz core. for a raw instruction rate of 8.4 gig or nearly twice as many instructions as the Power 6 per core. Now the x86 instruction set is not as efficient as the Power 6 but it more than beats that 4.7 number. Add to this a 2 socket, 4 core Opteron box has nearly double the memory bandwidth of the Power 6 does with two sockets and two cores and it really gets embarassing for the Power 6.

What is worse for IBM is that the Core 2 Duo is a four instruction per cycle machine with a top speed of 3.0 GHz making for a raw instruction rate of 12.0 gig or almost three times the Power 6 number. And it is far more efficient processing x86 so the number is much closer in real terms to the Power 6. So we have a processor that is faster than the Power 6 but is licensed at half the cost. If the Core 2 Duo memory system was any good IBM would be in terrible trouble, as it is we must wait until 2008 for Intel to fix that.

Tony

Those 120 Processor Value Units (PVUs) are pipelined and parallelized, resulting in a dramatic increase in value throughput by the Power6. Do you guys remember when each PVU was on a separate chip and most computers shipped with an empty PVU socket? My how times have changed!

For every $100 spent on IBM hardware they will charge you $## amount each year to use DB2 etc.

The number of instructions a chip can issue in a cycle is irrelevant. No general purpose CPU can sustain anywhere near its maximum issue rate, except in a very tight loop. It's impossible to predict the performance of a chip based on Instructions Per Cycle, number of threads or cores, or Mhz. The only way to make an intelligent decision is to measure it on a workload similar to the one you will be running.

I localize IBM software for a living. Compared to their documentation, this pricing scheme seems like the pinnacle of logic. And compared to their localization software, it sounds not just user-friendly, but actually will-give-user-a-bl**job-frinedly.

The term you are looking for is Confusopoly which is becoming a standard response by large corporations when they product they sell is a commodity with no discernable differences (worth noting anyway). As databases become a commodity, even given away for free, it becomes harder to differenciate on cost/performance/features. So just confuse the customer! This tactic is mainly used to retain current customers as opposed to gaining more, in that the customer can't easily see that the current product is compeditive or not. Witness mobile phone plans, Utilities etc to see who the confusopoly is taking hold!

http://en.wikipedia.org/wiki/Confusopoly

Per socket pricing is better?

All sockets are clearly not created equally.

What about the honking big MCM sockets on an IBM p595 which hold four POWER5+ chips (8 cores and 16 threads)? Should they be licensed the same as a socket holding one dual-core chip? Should socket doublers, like HP's mx2 module count as one or two sockets? What about modules like IBM's QCM or Intel's Clovertown?

One thing which is undeniable are cores (at least during this era of homogeneous cores). We know a Clovertown is four cores, but we can debate whether it is one chip or two (I would say two chips, one module).

We can also rate the performance of a core. An UltraSPARC T1 core is not designed to perform at the same level as a POWER6 core. So at the core level, they should not be licensed at the same cost. However, at the chip level, an UltraSPARC T1 produces about the same throughput for threaded commercial workloads (java, web, OLTP DB) as POWER6, so at the chip level the license costs should be similar.

A (number of cores) * (core performance factor) is the most fair way to license software. Which is why IBM and Oracle have adopted such methods.

But for fairness, the derivation of the core performance factor should be transparent and validated by a third party. IBM, as a processor and systems vendor, as well as a software vendor, has the power to manipulate software licenses to benefit their own servers.