Start Planning For New Systems Now

September 26, 2011 Timothy Prickett Morgan

Server buyers sure do have a lot to think about these days. Perhaps more than they have had to consider in a fourth quarter of a year in a number of years, in fact. That’s because there are a lot of new processors due from IBM, Intel, Advanced Micro Devices, and Oracle. The latter doesn’t have much of a direct effect on IBM i shops, right?

Correct–right up to the moment when Larry Ellison’s sales force suggests you dump your Power Systems gear and run JD Edwards EnterpriseOne on a cluster of Sparc T and Sparc Enterprise M systems. Trust me. It happens, and it will be happening more often as Oracle gets more aggressive about selling into the midrange market, trying to flip customers not only to Oracle iron, but Oracle databases and off DB2 for i.

It is hard to say who is going to get a new processor to market first, AMD or IBM. AMD is trying very hard to get its 16-core “Interlagos” Opteron 6200 processor, which is used in two-socket and four-socket X86 servers, to market as soon as it can to try to revive its flagging server market share. At the end of August, AMD had already started shipping these processors to its handful of vendors, and the good news is that they are socket-compatible with the prior generation of Opteron 6100 processors that debuted last March. The bad news is that because of Intel’s revival in the X86 server business, starting with the Xeon 5500s nearly three years ago, enthusiasm for Opteron-based servers have waned. The economy certainly didn’t help, in that server makers put their money behind Intel, more of a sure thing given where Intel’s engineering was at the time during the Great Recession. The result is that Sun Microsystems (now Oracle) and Fujitsu don’t sell Opteron machines any more, IBM has only one machine using the Opteron 6100, Dell and Hewlett-Packard have a pretty good lineup, and Acer, which is trying to break into the server racket, is stumbling along mostly not successfully.

The Opteron 6200 uses a new “Bulldozer core design that is interesting and is implemented in a 32 nanometer process that GlobalFoundries has had trouble ramping up, but it looks like the kinks have been worked out. The 16-core chip will deliver around 35 percent more performance on typical server applications compared to the existing 12-core Opteron 6100s, the company has said, which is the least you would expect from a chip that has been shrunk from 45 nanometers to 32 nanometers and has a new core design and 33 percent more cores at that.

The real question, and one that no one is yet able to answer, is will the Opteron 6200s compete well with the forthcoming “Sandy Bridge-EP” Xeon E5 series of processors for two-socket machines that will eventually be extended to four-socket boxes? Intel said at its Developer Forum that the chips have already begun shipping to big cloud and supercomputing customers, but that for marketing reasons it will not announce them until early next year. My guess is that the chip, which is the first X86 chip to include on-chip PCI-Express 3.0 peripheral controllers, is having some yield issues and Intel could sell all it could make to the cloud and supercomputing shops. So why make an announcement now and screw up sales of existing machines based on four-core and six-core Xeon 5600 processors? The server vendors don’t want that, and neither does Intel, with the PC market heading into a slump. AMD is not much of a threat with server makers knowing everything about the Xeon E5s and AMD not, as far as I can tell, offering the kind of compelling advantage it had in the X86 racket from 2003 through 2006.

Oracle is also launching the “Yosemite Falls” Sparc T4 processor today (September 26), which has eight cores with eight threads per core and which spans from one to four sockets in a single system image. This Sparc T4 processor is aimed directly at the belly of the Power Systems market for customers using the IBM i operating system in terms of performance on database workloads and also, unlike larger Sparc Enterprise M machines, supports Logical Domain (LDom) partitioning, which is pretty close to IBM’s PowerVM logical partitioning.

The Sparc T4 processors are different from the prior T series chips in ways that are going to make them more suitable for a wider range of workloads–the kinds that you find at midrange shops. The chips are expected to run at a higher clock speed, perhaps between 2.5 GHz and 3 GHz, maybe more if Oracle can squeeze it out of the 40 nanometer processes used to etch the chip by Oracle’s fab partner, Taiwan Semiconductor Manufacturing Corp. The prior two generations T2 and T3 chips ran at 1.2GHz to 1.65GHz and just didn’t have enough oomph for single-threaded workloads. In addition to more L3 cache and more dynamic thread control with the T4s, the new chips also have a little something Oracle calls the critical thread API that allows a single-threaded application to grab all the resources on a core and push that work through faster than a T3 chip. In early benchmark tests, Oracle said that the T4 has about the same throughput as the 16-core T3 chip on the TPC-C online transaction processing benchmark test, which stands to reason since the new T4 has half the cores running at about twice the speed. But Oracle also said that as gauged by the SPECint2000 single-threaded benchmark test, the T4 has nearly five times the performance as the T3 thanks to that critical thread API. (The question is how does the T4 stack up on these two tests against X86 and Power processors, of course.) The Sparc T4 chip also has improved on-chip cryptographic units that have between 1.5 and 3.5 times the performance of the Sparc T3 chips on encryption/decryption routings–like encrypting data in the database.

What does this mean to IBM i shops? There has never been a better time to have a set of Oracle, Intel, and AMD coffee mugs on your desk while inviting your IBM sales rep or business partner rep in for a little chat.

The Sparc T4 has IBM nervous enough to put out a big ol’ briefing for channel partners picking apart the chip and its systems before the details are even out. In that presentation, IBM points out that “Power7+ @ 32nm is just around the corner.”

Big Blue has not said much about the Power7+ chips publicly, and most of what it has said has been to The Four Hundred. Back in May, IBM was vague on whether the Power7+ chips would be socket compatible with the existing Power7 processors, and added that the Power7+ chips would have I/O upgrades that would require a system board swap. My guess is that IBM is goosing the GX bus for linking to remote I/O devices from DDR InfiniBand, which runs at 20 Gb/sec, to QDR or FDR InfiniBand, which run at 40 Gb/sec or 56 Gb/sec respectively. I also think that IBM will move to PCI-Express 3.0 peripheral slots and may even go so far as to put PCI-Express 3.0 controllers on the chip. IBM is transitioning from 45 nanometer to 32 nanometer processors in the hop from Power7 to Power7+ chips, and it has to do something. According to the updated Power roadmap I got ahold of a month ago, the Power7+ chips will have a “very large cache” and “accelerators” I also think IBM will goose the clock speed, maybe by 25 to 30 percent, but there is no guarantee that this will happen.

I think the announcement of the Power7 processor will be in October or November–and more precisely, I think it will be late October or early November. That gives IBM time to ramp up sales for the fourth quarter and to blunt the attacks from Intel, AMD, and Oracle in the midrange.

Here’s what you need to consider right now if you are shopping for a new Power Systems machine or an upgrade to one.

First, new products generally mean better bang for the buck. That doesn’t always mean a lower price, mind you. Sometimes, IBM gives you more bang for the same bucks to preserve its revenue stream. I happen to think this will be the case with the Power7+ chips, and depending on what IBM does on the I/O front, it might even charge a slight premium over Power7 processors if the I/O bandwidth is so large that overall system throughput skyrockets. The point is, if you need to buy a Power7 machine now and can’t wait for a Power7+ box, don’t let IBM get that impression and negotiate for a better price based on the fact that you think more bang for the buck is coming. Get a Power7+ briefing from IBM and find out more than I can.

Second, don’t just think about the hardware. Remember the software. IBM i shops pay for that operating system and database combo by the core. So for a given workload and CPW level, a Power7+ machine will require fewer cores to do a job than a Power7 machine. That’s a lower software bill, and software, not hardware, is the biggest part of the Power Systems-IBM i bill in the initial sticker price and over the long haul.