IBM Readies More Power8 Iron For Launch

July 28, 2014 Timothy Prickett Morgan

Big Blue is getting ready to flesh out the rest of the Power8 systems lineup, and it looks like the pace is picking up a bit. Back in June, IBM announced a four-core variant of the Power S814 single-socket server, a machine that was only tentatively expected to be launched in the third quarter and was not part of analyst or customer briefings. IBM had said very little about its plans for larger systems that gang up more than four Power8 processors.

In fact, the roadmaps that I have been able to get my hands on seemed to indicate that with the exception of this entry four-core machine–designed primarily for IBM i shops that do not need the extra processing capacity a full-on Power8 chip with six, eight, or 12 cores activated and who most certainly do need a machine in the lowest P05 software tier–there were no plans to get other Power8 iron out the door in 2014. This might be a too-literal reading of the Power Systems roadmap, but read what it says and you tell me:

To my way of interpreting American (as opposed to English), “rolling out over the next year and beyond” does not mean further machines based on Power8 chips should have been expected in 2014, but rather in 2015. But, people make mistakes when they make foils, and it could simply be a bad word choice and they meant to say “rolling out over this year and beyond.” It is hard to say.

But the impression I was given by sources at IBM as the entry Power8 machines were coming out was that IBM was focusing intently on scale-out clusters built from Power8 iron and not in any hurry to get scale-up machines out the door. But, IBM’s chief financial officer, Martin Schroeter, was pretty clear that more Power8 machines were indeed coming out this year. “We launched entry-level or scale-out Power8 in June, and had a good start compared to previous cycles,” Schroeter said in a call with Wall Street analysts going over Big Blue’s financials for the second quarter. “Keep in mind that entry-level is a small portion of the Power business. Power8 will be introduced into the midrange and high-end segments over the remainder of the year.”

So there you have it.

My suspicion is that yields were not great on the 22 nanometer process used to make the Power8 chips, and that is why IBM did not want to talk about the Power8 midrange and high-end machines potentially coming out this year when the entry machines were launched back in April. This idea is backed up by the fact that the initial Power8 machines actually use two half-chips of Power8s, not a full real Power8. This is called a dual-chip module, or DCM, and there is nothing wrong with it (with the possible exception that the cores in the two halves of the DCM can’t share a unified cache across the two half chips, as far as I know). IBM must have had similar yield issues on the 32 nanometer process it had to etch the Power7+ chips, because it used the same approach with these chips as well. To my way of thinking, IBM doesn’t have a high enough volume of Power chips to get the kinds of yields it needs to make big unified processors with a dozen cores at a price that it can afford.

Presumably, IBM is ramping up the yields on the Power8 chips and does not have to resort to such DCMs, but maybe not. If that was the case, then perhaps IBM should have designed a six-core Power8 from the get-go and made them explicitly to connect instead of having to cut dud parts out of the chips. The Power7+ versions of the Power 750+ and Power760+ used these DCMs, and it is possible that future four-socket machine, presumably called a Power S845 because it is a scale-out machine with four sockets and in a 5U enclosure will also use DCMs. The Power7 machines had four cores in a 4U enclosure, while the Power7+ had four sockets in a 5U enclosure for the 750-class box. The Power760+ was the same machine, but equipped with DCMs with a higher core count.

If a Power 770+ type of machine emerges, it would probably use a similar interconnect and use the expanded on-chip NUMA clustering to put 16 sockets in a single system image. This machine would probably be built from a cluster of the four-socket, 5U boxes because density is less of a priority in these larger machines and local disk capacity expansion is. So this box would have four nodes with a total of 16 sockets, which makes it an S81620 if the naming convention holds. But, IBM could drop the S for such a machine because it is not scale out, but scale up–meaning having a single shared memory space using NUMA across those sockets.

IBM has not given much in the way of any details about how it will replace the current Power 795, which is now four years old and looking pretty long in the tooth with its Power7 processors. IBM could no doubt build a very large NUMA machine, perhaps with 32 or 64 sockets and anywhere from 384 to 768 cores behind perhaps as much as 32 TB or 64 TB of main memory.

What IBM does will depend in large measure on what Power Systems customers are asking for in a scale-up machine–if anything at all. It could be that IBM can address most of the needs of the customer base with a 16-socket machine. Two 32-core partitions on a top-end Power 795, each using 4 GHz Power7 processors with four threads per core, can push 399,200 on the Commercial Workload Performance benchmark test that IBM uses to gauge relative performance on IBM i workloads. So a full-on 256-core machine should, in theory, rate at around 1.6 million CPWs if IBM i scaled perfectly across that machine. (Which is does not.) A Power8 machine with sixteen sockets but 50 percent more cores yields somewhere between 2X and 2.5X the performance per socket at a nominal 4 GHz clock speed. Call it 2X to be simple. Such a Power8 machine with all of its cores activated would have 192 cores and deliver something like 2.5 million CPWs.

If IBM needs anything, it needs fat memory four-socket and eight-socket machines to better compete against Intel’s Xeon E7 processors, machines that have high core count as well as big gobs of main memory. If Oracle is really selling the Sparc M6-32 machines, which have 32 sockets and support 32 TB of memory, then it needs a box in this class as well. It would also be good if IBM made some tweaks to IBM i pricing for software tiers P10 through P50, but I am always asking for the software to be less costly. That is because this is where the real money gets spent, and you notice that IBM never lowers the price there, right? And with no clone of IBM i, there is no real alternative to do so.

We shall see what IBM does. But I suspect that excepting a few hundred very large IBM i shops, the very big iron is important only to AIX shops and the midrange four-socket machines will be much more important to those running IBM i. Even if customers don’t need the oomph of such a four-socket machine, they want to know it is available as their applications and databases grow beyond the capability of the two-socket Power8 machines. I would not be surprised to see two different styles of four-socket and maybe even eight-socket machines, which have skinny and fat main memory configurations and maybe differing amounts of local disk and flash storage.