It is three years later, and now IBM wants to do it again with the Power5-based Squadron p5 590 and 595 servers, which will very likely be the performance and price/performance leaders in the server market for quite some time. IBM hasn’t announced TPC-C online transaction processing benchmarks for the new 64-way p5 595s, but the word on the street is that it might be able to break the 3 million transactions per minute barrier. That’s three times what IBM could deliver with its 32-way pSeries 690 machines, which used the 1.9GHz Power4+ chips, and it is also three times what HP has been able to show with its 64-way Integrity Superdome servers, which use Intel’s 1.5GHz/6MB cache Itanium 2 processors.
Even doubling up on the Itanium 2s with the mx2 dual-Madison modules that HP invented itself, and moving to a 1.8GHz/9MB cache Itanium 2 chip (which should have been here by now), HP might be able to break 2.2 million TPM on the TPC-C test. And with 144 cores running at 1.2GHz, Sun would probably be able to show maybe only 1.5 million TPM. That is a lot of open water for IBM to sail in during 2005 and into 2006, when HP and Sun might be able to field more performance in the enterprise server space.
Karl Freund, vice president of product marketing for IBM’s pSeries Division, says it is really quite simple. With the Regatta, we had great performance. But the Squadron line is a game changer. We believe that these machines offer us another inflection point in our growth, and our goal is nothing other than to get to a 40% revenue market share in the Unix server business. Maybe IBM should have nicknamed the Power5 servers Gauntlet. Not only has IBM thrown one down at both HP’s and Sun’s feet, but now both HP and Sun are going to have to run one each time IBM tries to sell into their accounts.
The new Squadron servers complete the Power5 product roll out that began with the eServer i5 OS/400-based servers in May and continued with the eServer p5 AIX-based servers in August. All of these servers are based on the dual-core Power5 chips, which are available at 1.5GHz, 1.65GHz, and 1.9GHz clock speeds. IBM is already shipping a two-way p5 520, a four-way p5 550, and a p5 570 that scales from four to 16 processors.
IBM is selling the top-end Squadron servers in two variations. The p5 590 uses the 1.65GHz Power5 cores and supports from eight to 32 processors in a single system image. These appear to be based on the same multichip module approach used in the pSeries 690, which took four Power4 cores, glued them into a single MCM, and then put four of these in a machine to make a 32-way. This machine offers from 8GB to 1TB of main memory (only 512GB is available until double-density memory cards are available, in the second quarter of 2005) and up to 9.3TB of internal disk capacity. The machine has 160 PCI-X slots, and it occupies a rack and a half of space. An eight-way p5 590 with 16GB of main memory and eight 73 GB disks costs $451,000; a 16-way with 32GB of main memory and 16 73GB disks costs $745,000, and a 32-way with 64GB of main memory and 16 disks costs $1.42m.
The p5 595 doubles up the number of cores on a Squadron cell board from eight to 16 (that’s two MCMs, not a double dense MCM). The bigger p5 595 spans from 16 to 64 processor cores. Either 1.65GHz or 1.9GHz Power5 cores can be used to build a system. This server has a main memory that spans from 8GB to 1TB today, and again next year that memory can be doubled to 2TB. This server has 240 PCI-X slots and can fit 14 TB of disk capacity internally. The 16-way version of the p5 595 using 1.65GHz cores sells for $904,000 with 32GB of main memory and 16 disk drives; while a base 64-way system will sell for a little over $4m, according to IBM. Generally speaking, the p5 Squadron boxes using the 1.65GHz Power5 cores offer about 45% more performance than the Regatta boxes using the 1.9GHz Power4+ cores in a like-for-like configuration, and cost 40% less. This is a significant improvement in bang for the buck.
Like the other Squadron boxes, the p5 590 and 595 servers run AIX 5L 5.2, the release of IBM’s Unix that is also supported on the Regatta machines. This release offers limited logical partitioning capabilities but has the virtue of having over 4,000 applications certified to run on it, according to Freund. AIX 5L 5.3 is necessary to support the Virtualization Engine features, such as virtual LAN and I/O and micropartitioning that allows up to 10 partitions per physical processor core (up to a maximum of 254 per system, which is a tested limit against the maximum theoretical limit of 640 partitions on a 64-way). It took IBM about 15 months to get those 4,000 applications certified on AIX 5.2, and it will probably take as long to do it on AIX 5.3 as well. Hence, IBM supports AIX 5.2 on the Squadrons, even though in the past it might have required AIX 5.3. AIX 5.3 was about a year late to market, and it can ill afford to make it a condition of a server purchase.
One thing the press releases don’t talk about that makes a big difference with the Squadrons is simultaneous multithreading. The Power5 chips have not only two cores per chip but also two threads per core. This SMT support is built into the hardware and woven into the AIX 5.3 kernel and does not require a recompilation for it to take effect, according to Freund. Most significantly, SMT provides anywhere from 30% to 45% performance boost on multithreaded applications, depending on the code. If Intel didn’t think HyperThreading matched well with multiple cores, you need only look at the Power5 to see that maybe Intel is wrong.
