Big Blue’s Power Systems Get Some Itanium Competition
November 12, 2012 Timothy Prickett Morgan
It hasn’t been an easy couple of years for IBM in the RISC/Unix midrange and enterprise server market. And it has just gotten a bit harder and stands to get even more difficult in the next year. Not only did Intel finally get its “Poulson” Itanium 9500 processors into the field, which will be picked up by Hewlett-Packard for its HP-UX, OpenVMS, and NonStop systems, but it made a broader commitment to converge the high-end of the Xeon processor line with the Itanium line with the next generation “Kittson” processors.
And Oracle is working on its next-generation Sparc T5 processors for entry and midrange machines and its own Sparc M4 processors for midrange and high-end servers, due respectively in the spring and fall of 2013.
In other words, the easy pickings among HP-UX, OpenVMS, and Solaris shops for IBM’s Power Systems division is over. And that impacts your life as a Power Systems shop in a number of different ways. IBM will have to work hard to get you to invest in more Power iron and systems software to make up the difference as Sparc and Itanium shops breathe a little more easily than perhaps they have in the past three years.
The situation is complicated, of course, and is not just solved by chucking a bunch of processors into the market. Since March 2011, Oracle has been waging a war against the Itanium platform, saying that Intel was going to mothball it and that HP knew as much and was not being honest with its customers about it. And thus, Oracle decided then and there that it would not support its future releases of software on the Poulson and Kittson Itaniums. HP and Oracle came to legal blows, and in August Judge James Kleinberg, ruling from Santa Clara County Superior Court in California, ruled that Oracle had a contractual obligation to keep porting and maintaining its code on Itanium processors, and in September, the company said it had restarted this work to comply with the court order even as it seeks to appeal the ruling. Here we are in November, and the new Poulson Itanium 9500 processors are out, and HP is getting ready to ship entry Integrity blade and rack servers using it in December and high-end Superdome 2 servers using it in January.
While HP has done some work to tweak its Integrity and Superdome 2 machines to absorb the Itanium 9500 chips, these boxes were designed from the get-go to use the eight-core Itanium 9500s as well as the quad-core “Tukwila” Itanium 9300s that launched in March 2010 and made their way into HP machines the following April. The important thing as far as HP-UX customers are concerned, who tend to use Oracle databases and middleware as well as Oracle and SAP application software, is that the new Poulson-based machines use the exact same HP-UX 11i v3 operating system as the Tukwila boxes. HP snuck out an update of HP-UX 11i v3 in September that had patches for the Poulsons, but because the runtimes in the operating system did not change, by default any software that ran on HP-UX on Itanium 9300s will run on Itanium 9500s. And you don’t have to recompile applications to see an extraordinary boost in performance. Which is a good thing because I highly doubt that Oracle will be taking advantage of those Poulson features to recompile its database, middleware, application software. (The court said Oracle had to support Itanium. It didn’t say how well Oracle had to support Itanium.)
The Poulson chips pack a number of punches that HP is going to be able to leverage to stability its Itanium-based server businesses. First, you have double the cores on the Itanium 9500 die. Then you have higher clock speeds and lower thermals at the same time. The top-end four-core Itanium 9350 ran at 1.73 GHz, had 24 MB of L3 cache, and burned 185 watts. The top-bin Itanium 9560 has eight cores, which run at 2.53 GHz, plus 32 MB of L3 cache; it only burns at 170 watts, thanks to a shrink to 32 nanometer processes. The Poulson chips have two half QuickPath Interconnect (QPI) point to point links plus the full four speed QPI links to lash the processors to each other and to peripherals in the system. Those QPI links run at 6.4 GT/sec (giga transfers per second), which is 33 percent faster than the Itanium 9300s had. In the labs, Intel says that the Itanium 9500s offer anywhere from 2X to 2.5X the performance, socket for socket, as the Itanium 9300s in the same relative SKU in the lineup. And that is without recompiling applications to take advantage of a new pipeline that can issue twelve instructions per clock instead of the six of the Tukwila pipeline.
In its Integrity and Superdome 2 systems, HP is telling customers running HP-UX that they can see as much as 3X to 3.7X the performance moving from Tukwila to Poulson processors. Even if you spread that performance jump over nearly three years, it is still a very large performance jump. And here is the very important thing. Ric Lewis, the new general manager of HP’s Business Critical Systems division, told customers at the Itanium 9500 launch event in San Francisco that HP would keep its hardware prices the same for Poulson-based Integrity and Superdome 2 machines. And here is the real cool bit for HP-UX customers: HP-UX is priced on a per-socket basis, not on a per user basis, so they are not going to get shellacked by a huge software bill like you will if you try to move from a Power6 to a Power7 or Power7+ system, where IBM charges per core for operating systems and per processor value unit (PVU) per core on other software.
It is not entirely clear how OpenVMS or NonStop are going to be priced on the new systems, but you can bet that HP is going to do as little as possible to antagonize these shops. After seeing Integrity and Superdome 2 sales crash thanks the fight over Itanium’s fate with Oracle, all that HP wants to do is calm customers down and get them buying again.
What will also be calming to HP shops is the fact that after a decade of making hints about a common socket for Itanium and Xeon, it looks like this will happen with the future “Kittson” Itanium chip, due in two to three years from now. The idea, Intel explained last week, was to use common blocks of transistors for cache memory, main memory controllers, and other components of the chips across both Xeons and Itaniums while retaining the unique Xeon and Itanium cores and their instruction sets. (IBM has been doing this for Power and System z chips for years.) Intel will also move to common packaging and sockets for Itanium and Xeon E7 server processors, which will mean server makers like HP can make a Xeon E7 server that can also snap in an Itanium processor without having to change anything else about the system.
My guess is that the future Xeon E7, most likely in the “Haswell” generation due in 2014 at the low end and maybe 2015 at the high end, will actually plug into the current Itanium socket. Intel had promised two and a half years ago when the Itanium 9300s launched (and very late for that matter) that Tukwila, Poulson, and Kittson would all share the same socket. This is the only way to make both sets of statements true.
IBM has to be thinking the same thing about future Power processors. There has to be some way to converge X86, Power, and mainframe processors in some way. It would radically simplify IBM’s server lineup, and make for less hassles. Just for fun, imagine this: IBM buys Advanced Micro Devices, converges Power, Opteron, and z engines to Opteron sockets, and gets into the ARM server racket to boot now that AMD is an ARM licensee and is looking to get 64-bit ARM server processors out the door by 2014.
Wouldn’t that be fun? Probably won’t happen though. That said, there has never been a better time to buy AMD.