Robert Wessel <[email protected]> writes: > My entire point was that the *smallest* systems have the best price > per processor chip (although that's entirely meaningless for the > customer, since customer don't buy CPU chips, they buy performance, > which does get better on a per-dollar basis on larger systems). > > Consider my example, a single-CP, kneecapped* EC12 (a 2827-401) ships > with a single book containing six CPU chips, and nominally should cost > about $840K. *Exactly* the same physical system (with the same > processor book containing the same six CPU chips), with the magic > license codes installed to turn it into a 20-CP full-speed EC12 (a > 2827-720), sells for $10.8M. In the case of the -401, IBM is > effectively selling CPU chips for $140K each (assuming that's the only > thing in the box), but in the case of the -720, IBM is selling them > for $1.8M each. IOW, if IBM sold only the smaller systems (as opposed > to only the karger systems), the same revenue would result in 13 times > more CPU chips physically shipped to customers. > > If you wanted a -721 (one more CP than the -720), you have to get a > *two* book system, and you'd pay $11.2M and get a dozen CPU chips in > the box (sending the price-per-CPU-chip back down to $933K). > > And if any customers buy IFLs, those are about $55K each on an EC12, > and zIIPs and zAAPs are about $100K each. > > A maxed out EC12 physically has 144 CPU cores on 26 six-core chips. > Only 120 of them are active, and only 101 of them can be used for > customer workload. > > > *a 2827-4xx runs at about the sixth of the speed of a -7xx box, so a > single CP 401 gets about 240 MIPS, while a (single CP) -701 does 1514.
a decade-plus ago, i did some work for financial processor ... that did max'ed out mainframe systems and rolled them over every generation. at the time, their largest datacenter had 43 such systems ... but there were a number of others in the financial industry that had larger operations. in aggregate those institutions accounted for majority of ibm mainframe processor revenue. since ibm hasn't given details of their market, i've used annual processor revenue divided by max. system price to provide number of max. mainframe equivalents per year (since that is also the majority of the revenue). part of ibm's dilemma has been that mainframe system throughput has tended to increase faster than those institutions mainframe workload increases. the particular 43 system complex was based on a single application that ran every night and needed to finish in the overnight batch window ... even tho it had a large performance group provided for its care & feeding ... i was able to find another 14%. I've periodically wondered if ibm mainframe pricing is somewhat along the lines of airline seat sales ... if you already flying the plane ... then getting any money at all for otherwise empty seats is better than nothing. when i did chips, the smallest wafer run i typically could get was six wafers. in 14nm technology & 450mm wafers yielding 1383 chips, a minimum run would be 8298 chips. the major customer base currently is maybe 1100-1200/annum ... then there would be a significant number of left-over chips that need to be unloaded before the introduction of the next generation (chips). And just like airline seats ... they wouldn't want to see those paying the premium prices switching to significantly discounted product. trivia: the financial processor had been spun off from larger financial institution (in 1992, in the largest IPO up until that time) that guerstner had been president of. some speculation that ibm board bringing in Gerstner was at the behest of this financial market segment to resurrect ibm (mainframe) and reverse the breakup (into the 13 baby blues). ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
