What Price Power (Eight And Maybe Nine)?
July 11, 2016 Timothy Prickett Morgan
Call me old school if you want, but I believe in Moore’s Law and I believe that IT vendors have to keep giving customers more bang for the buck if they want organizations to keep investing in technology. There are basically only two levers to help push a new technology into the market, and that is increasing the performance of a device or lowering its price, the latter hopefully occurring if the cost of production comes down but sometimes not as vendors seek to maximize their profits and make it up with performance leaps.
This has certainly been the case with the central processors used in servers for the past several generations of chips, be they from Intel or IBM. In general, chip makers have tried to keep their CPU prices relatively constant and giving some of the Moore’s Law advancement to users with increased performance (sometimes per core, sometimes across more cores) while pocketing the rest as profits.
As you all know, IBM did not put out the Power8+ chip enhancement to the Power8 processor that many had expected it to deliver this year, and it has now been 27 months since the Power8 chips debuted and–guess what?–the prices on the Power Systems S-class machines that support IBM i (in addition to Linux and AIX) have not changed. Not by a single penny.
This violates my sensibilities, and it probably violates yours, too.
This is particularly true when you consider that during the Power8 generation, Intel has launched the “Haswell” and “Broadwell” Xeon E5 v3 and v4 processors and moved more or less down the Moore’s Law curve to offer customers better value for the compute dollar. When the Haswell Xeon E5s launched in September 2014, five months after the Power8 chips first debuted in the S-class Power Systems, the earlier “Ivy Bridge” Xeon E5 v2 processors topped out at 12 cores. The Haswells scaled up to 18 cores and offered about a 40 percent performance bump, on average across many workloads, and the Broadwells scale up to 22 cores and offer another 27 percent more oomph on average. Call it a 70 percent performance jump over the same timeframe that the Power8 has been in the field. When the Power8 launched, base systems offered about 25 percent to 30 percent more performance at about the same price over a base Power7+ configuration. In the case of the Intel boxes, the big difference in the price/performance at the processor level is muted to a certain extent by the other components in the system, which change their capacities and prices at their own individual rates.
For the sake of argument, let’s say at the system level and not using top-bin parts but rather more standard configurations with fewer cores (and lower prices per unit of compute) that at the system level the Xeon server platform has improved its performance by 70 percent (this is for E5-2640 parts), and the price of the processor has risen by about 6 percent and the system price has probably risen by maybe 10 percent with richer configurations including more memory and flash storage. If you do the math, at the system level for a base machine with 20 cores, the bang for the buck has improved by around 35 percent over those ensuing years.
To keep its relative competitive position in the market, therefore, it stands to reason that Power Systems machines have to deliver a similar 35 percent improvement in price/performance. I doubt very much IBM could have done this without a lot of effort with the Power8+ chips, had they materialized, and it will have to do even better when the Power9 chips come out next year because by then Intel will have rolled out its “Skylake” Xeon E5 v5 processors, which will probably have about 25 to 30 percent more performance per chip than the Broadwells from a mix of per-core enhancements and more cores on the die. Call it twice the performance of the Ivy Bridge Xeon E5 v2 chips from 2013 at a system price that is maybe 15 percent higher, so that is around a 45 percent improvement in price/performance over the same span from Power7+ to what should have been Power8+ and what should be the target for Power9 now.
Without a yank of the performance lever for the Power8+ chips that did not happen this year, the only lever to pull on is price. So what does a 35 percent price/performance improvement look like? Take a look at what the prices for Power8 iron in 2016 should look like:
Back in the day, when a new generation of AS/400 and iSeries machines came into the field, the new machines immediately sold for at least 15 percent off list price, and the price on the older gear was pushed down proportionately to the relative performance it offered compared to the new gear. It wasn’t too long before a new generation of iron came out, and this pushed the gear in the prior two generations down so far that you could get an N-2 generation system for 50 percent off list and an N-1 generation system for about 35 percent off list price. The performance increases were relatively small, but predictable, and over time the hardware prices kept coming down as IBM shared the Moore’s Law benefits with customers.
Discounting for old or new machines has pretty much all but stopped, and that is because the margins for Power Systems-IBM i platform resellers are so skinny that they can’t actually afford to do it. According to the dealers that I have spoken with, there are some special add-on deals that can sweeten a new system deal, but not much in the way of discounting as we knew it then. The reason is simple: IBM i shops are either going to buy a new system, or they will delay an upgrade or move off the platform. These dealers tell me that if IBM cut prices on the Power8 servers now, all it would do is reduce their revenues and IBM’s along with it, because customers will either leave or stay. One dealer told me that he reckons based on his customer base that about 10 percent of the IBM i base is going to leave and another 30 percent are threatening to, but the other 60 percent are not going anywhere because they have too much invested in their application software. Many of these shops are sitting on old i5/OS or IBM i releases and they have little inclination to move because the fees to upgrade their software far outscale any potential decreases in hardware costs that IBM might provide.
I countered that what IBM should then do is cut the prices and offer beefier configurations that encourage customers to move more workloads onto the IBM i platform, perhaps adding Linux workloads, too. This would keep the revenue stream high and the partner margins high, but it would eat into IBM’s margins because it was putting more hardware in the box for the same price. Big Blue is not going to do that, either.
And so here we sit in a base that is not able to consume capacity like the X86 server market. This is a real problem, far more troublesome than any point-in-time comparison between Xeon and Power platforms.
If we look ahead to what the Power9 systems might look like, I think the per-core performance will rise by about 27 percent based on some back-of-the-envelope math. If you use the pricing levels I suggest for 2016 on machines with the same number of cores as the Power8 boxes, then lo and behold you give customers precisely 2X more bang for the buck on IBM i workloads for the jump from Power8 and Power9. And, IBM will have the capability to offer twice as many cores per box, too, and dial up its revenues by stuffing the box even more full of cores.
The trick, it seems to me, is to get customers to do a lot more work on the Power Systems machines. The capacity is going to be there, for sure. I will consider how to do this, and what effects it might have on the IBM i market, in a future issue. I don’t like any answer that does not help IBM and its partners make money while at the same time growing the aggregate compute capacity used by the IBM i market at something at least approximating Moore’s Law improvements in pricing.