Let’s not pull any punches. The G5 series of S/390 servers, which IBM announced today (see Top Stories section), are by far the most impressive mainframes that IBM has offered to its enterprise System/360, System/370 or System/390 customers. The machines not only sport the fastest mainframe processors ever developed by IBM, but also the increased memory and I/O bandwidth capacities to put them in the same class as the fastest Unix boxes in the market. IBM may be talking about how the G5s will allow the company to expand its base of mainframe customers and get existing customers to use S/390s to support new workloads such as ERP applications from SAP, PeopleSoft, Oracle or Baan as well as supporting advanced applications such as data warehousing and e- business extensions to legacy applications, but Hitachi Data Systems and its Skyline series of bipolar mainframes are what IBM is attempting to shoot down with G5. The Turbo Symphonies, which clock at 465 megahertz and provide over 125 MIPS of processing power, are more than fast enough to replace all but the largest Hitachi Skyline processors, which offer 150 MIPS of power per engine. Over the last few years, IBM’s CMOS mainframes had processors that ran at only 45 and then 63 MIPS, and for many big banks, insurance companies and manufacturers, this was not enough power. That’s why Hitachi has sold hundreds of Skylines since IBM left the bipolar fray.

By Timothy Prickett Morgan

Skyline has been taking it out of us over the past few years, concedes Bill Zeitler, IBM general manager of Server Brand Management. And we are going to get it back with G5. According to Hitachi’s own surveys of the S/390 installed base, about 4 percent of the base consumes 40 percent of the MIPS each year, and thanks to Skyline, Hitachi has been able to get half of that MIPS consumption rather than the 5 to 10 percent that has been its traditional market share in the mainframe market. By all measures, Skyline has been a phenomenal success. But G5 changes everything, even more than Skyline or the first four generations of IBM CMOS mainframes. The big names on Wall Street are pretty much in consensus that IBM has got Hitachi squarely in its crosshairs with the G5 S/390s. To the dismay of Hitachi Data Systems, it sounds like G5 performance is coming in better than expected, says Stephen Smith, managing director and IBM analyst at PaineWebber. If closing the gap in performance with Hitachi Data Systems, which is selling Skylines at $8,000 per MIPS, means that IBM could hold pricing close to the $6,000 range for a year, then it has a shot at some growth and this could also be positive for margins. Still, IBM is really challenged to grow this business for any meaningful period. (It may not sound like it, but that’s high praise from Smith, who has been rightfully skeptical of IBM’s server and PC businesses for as long as anyone can remember). Steven Milunovich, first vice president at Merrill Lynch and the lead IBM analyst at the company, is similarly upbeat about IBM’s prospects. Milunovich last year replaced long- time IBM analyst Dan Mandresh, who many years ago shared office space with Ulric Weil and Gideon Gartner, two other extremely influential ex-IBMers. Mandresh was perhaps the most significant stabilizing factor on Wall Street that helped investors keep level heads about IBM’s stock through the dark years before and just after Louis Gerstner became chairman and CEO of IBM in 1993. Milunovich is, for the short term, optimistic about how G5 will play in the mainframe market. In a May 5 stock report, Milunovich rated IBM a long term buy and a short-term accumulate, and expects the announcements to help push IBM stock up to $125 from its current $118. The Symphony processor is one of the most complicated pieces of electronics that IBM has ever built and shows that IBM has been able to adapt (however belatedly) to the demands of a faster-paced server market. It used to take IBM three years to get a new generation of bipolar mainframes out the door (for economic as well as technical reasons, mostly economic), but it has introduced five lines of CMOS mainframes in three and a half years. The 32-bit Symphony processor uses IBM’s CMOS 6X process technology and has over 25 million transistors in it; the chip’s L2 cache memory has over twice as many transistors. (Don’t expect to see the CMOS 7S copper chip process being used in S/390s until next year.) These chips also include support for IEEE decimal floating point calculations – the kind used by Unix applications – as well as the hexadecimal floating point traditionally used by IBM mainframes. While this seems like a trivial change, many Unix applications – Lotus Domino and various ERP applications come immediately to mind – make heavy use of decimal FP and this support will significantly improve the performance of these applications when they are ported to the OS/390 Unix shell to run on S/390 servers. The Turbo Symphonies may offer 125 MIPS, but the standard Symphony chips used in most of the 15 new 9672 models announced on May 7 will provide 115 MIPS of power. The models using the slower G5 chips will ship in the third week of August, while the Turbo models will ship in mid-September. IBM will begin early shipments to selected customers in July.

Multichip module

A G5 multichip module (MCM), the core of every new 9672 server, includes twelve of these processors plus two integrated cryptographic processors. Ten of the G5 chips can be used as CPUs in an SMP single image system; the two others in the MCM can be either used as hot backup in the event of a chip failure or as 3990-style disk controller (IBM ported the 3990 controller program to the S/390 CMOS chip family during the G4 generation). Each G5 chip has twice the power of IBM’s H5 bipolar engine (which was used in 9021 mainframes), and the G5 MCM used one- fifth the electricity and one-tenth the glue chips as were required in the H5’s bipolar thermal conduction module (TCM). With ten processors activated in the MCM, a G5 server yields 900 MIPS, which is twice the capacity of the ten-way 9672-RY5 server and nearly twice power of the ten-way 9021-9X2 bipolar system IBM shipped in early 1994 as a stopgap measure to sell against Skyline until the G4s and G5s became available. IBM says that it has improved its SMP ratios, but that is not, strictly speaking, true. As processor clock speeds get cranked up, SMP ratios usually degrade considerably – there’s just a whole lot less time to juggle instructions and data around between clocks; IBM has held its SMP ratios while increasing clock speeds by increasing L2 cache memories and by getting rid of cache contention that it could not work the kinks out of in the G4 series. The plan for G5 servers was initially to deliver a ten-way with 650 to 700 MIPS, so 900 is looking pretty good to IBM. Early next year, IBM will reportedly ship a twelve-way G5 machine that will offer about 1,200 MIPS. No one is talking about G6 yet – those will use copper chip technology – but it wouldn’t be outrageous to see a 20% to 30% performance increase just from moving to the new chip process alone. That will put IBM’s CMOS on par with Skyline, and put the heat on Hitachi to either get moving with Skyline2 bipolar processors, which could have as much as 225 to 250 MIPS of power, or concede defeat and get moving trying to peddle its own Pilot CMOS mainframes, which, by the way, use IBM G series chips. Of course, processors are not the only factor when it comes to overall server performance, and IBM has been careful to balance all the other resources of the G5 series so customers can make use of all this power. The G5 servers can support 24 gigabytes of main memory, up from 16 gigabytes with the G4 servers and 8 gigabytes with the 9021s. The G5s sport four internal I/O buses, which provide double the I/O bandwidth of the G4 servers, and will also include a new Fibre Channel peripheral connection (FICON) scheme that is a significant improvement over IBM’s ESCON fiber optics. FICON will offer 100 megabyte per second channel speeds compared to the 17 megabyte per second ESCON speeds; initial G5 shipments will reportedly offer 12 of these fiber links, increasing to 36 next year. While most mainframes offer up to 256 ESCON channels, according to Merrill Lynch, the average 9672 shipping today has only 20 channels.

Faster sysplex

IBM has other tricks up its sleeve, of course, to try to get its mainframe base excited about G5 machines and the G4 models that they will have to accept today as they wait for G5s to start shipping in the fall. For big customers who use its parallel sysplex clusters to create scalable servers, IBM has created a faster sysplex based on what is called the Integrated Cluster Bus, which provides 280 megabyte per second links between clustered systems, up from 100 megabytes per second with the prior generation of parallel sysplex. IBM is also introducing a modified sysplex called a parallel geoplex, which will allow companies to cluster machines that are as far as 40 kilometers apart; this, says IBM, is a requirement that its customers want to ensure a disaster doesn’t take out all their mainframes in one fell swoop. IBM is also cutting software prices in an effort to get customers moving not only to new hardware, but up from VM and MVS to OS/390 as well as to new data base and middleware programs. For quite a while now, IBM has been offering customers discounts if they move to parallel sysplex software licensing (PSLC); this was a way to get customers to at least install the sysplex even if they didn’t use it. (Late last year, IBM reportedly started requiring that customers push at least 50 percent of their transactions through the sysplex; they couldn’t install the sysplex software just to get the price break.) The PSLC licensing saves many mainframe customers as much as 25 to 30 percent over what they would pay for non-PSLC licenses; but IBM says its biggest mainframe customers – those who just paid a bundle to run OS/390 on Skylines, for instance – would only see 5 to 10 percent software savings if they went to parallel sysplex. That’s why IBM is chopping software prices by 25 percent for all S/390 customers with more than 1,000 aggregate MIPS in their IBM systems. IBM is also moving from capacity-based pricing on its CICS, MQSeries, IMS and other middleware programs for mainframes to a straight user-based pricing scheme; it hasn’t announced what those user-based prices will be when they are announced in September and early next year, but in the interim IBM is chopping capacity-based fees by 10 percent as a gesture of good faith. IBM says that it has been bidding G4 placeholders as part of G5 deals since January and that it has hundreds of G5 customers lined up already. IBM is adamant about not talking pricing, even though it is required to by law. The 1956 Consent Decree, which IBM signed to put a stop to an antitrust lawsuit brought against it by the US Department of Justice, is still in effect for the AS/400 and S/390 lines and requires IBM to release list prices for all of the equipment and software it sells under these lines. (Bill Gates is not the only bad boy in business.) Expect prices to be around $6,000 per MIPS for complete G5 systems, down from $7,000 per MIPS for G4 servers six months ago.

á