Fujitsu’s future Olympus processor will be at the heart of the so-called Advanced Product Line (APL) servers that Fujitsu, Siemens, and Sun Microsystems will begin selling in the middle of 2006 or so. Olympus is a variation of the Sparc64 VI processor that Fujitsu was designing for its PrimePower Unix servers well before Sun finally came to its senses in June 2004 and partnered with Fujitsu to sell a mixed (though not merged) UltraSparc and Sparc64 Unix server line.
Exactly what makes the Olympus chip, which has some design input from Sun, different from the Sparc64 VI processor that Fujitsu was already designing is unclear, and neither Sun nor Fujitsu have tried to clear that up. Many observers are beginning to wonder if Sun is doing anything in the Olympus design beyond what it would have been doing anyway, which is porting Solaris 10 to it, as it has ported Solaris 8 and Solaris 9 to past Fujitsu processors on behalf of Fujitsu-Siemens.
Richard McCormack, vice president of product and solutions marketing for the partnership’s Fujitsu Computer Systems unit in North America, says that the Olympus-based APL products – there are other APL servers that will be based on Sun’s future Niagara multithreaded Sparc chips – are right on schedule for shipments some time in the middle of 2006. The code name for servers based on the Olympus chips is Jupiter, if you like to play code name bingo. In fact, Solaris 10 is being beta tested on APL machines right now at limited customer sites, according to Mr McCormack.
By the way, Mr McCormack says that the APL partnership does not include any of Sun’s current Opteron-based servers – the V20z and the V40z – and it will not include the future Galaxy Opteron machines that are being created by Sun founder Andy Bechtolsheim.
Sun was expected to start its launch of the initial Galaxy servers in January or February of this year, but the products have been pushed out. Sun is not saying why, but delays are to be expected with any radical departure in design. Fujitsu-Siemens has not, to his knowledge, sold any Sun Enterprise UltraSparc-based servers to date in North America, but the company has a long-standing OEM agreement with Sun for the Japanese market. Sources at Sun say that it has indeed sold some PrimePower iron under the APL alliance in the United States.
Mr McCormack says further that Fujitsu-Siemens has been selling Sun iron, such as its DC-powered, NEBS-compliant UltraSparc servers for telecom companies and service providers. The Sun and Fujitsu-Siemens sales forces remain separate, and we do compete with each other, he adds.
The Jupiter servers are so-called flat SMP servers, which have high-speed buses that link cell boards into a single system image for an operating system – in this case Solaris – to run its applications on. The Olympus processor is most likely a tweaked version of the Sparc64 VI, which was originally due for the PrimePower servers in late 2005 or early 2006 according to roadmaps from two years ago.
Like Olympus, the Sparc64-VI was to be implemented in a 90 nanometer copper/SOI process and made in Fujitsu’s labs in Japan; it was to have 128KB of on-chip L1 data/instruction cache per core, 6MB of on-chip L2 cache memory shared by both cores (with perhaps a 4MB cache variant), and an initial target speed of 2.4GHz, with speeds eventually ranging between 2.1GHz and 2.6GHz. Back in June, Sun said that APL machines would span to 64-way SMP, which seemed to imply that Fujitsu-Siemens was expected to back off from the top-end 128-way processing it already offers in the PrimePower line.
With Sun and Fujitsu already having worked to get Solaris to scale on 128-way SMP in the PrimePowers, it seems unlikely that the two companies would push beyond this limit. What APL machines might do is have 64 processor sockets, and given that the Olympus chips have two cores each, the Jupiter server can scale to 128-way SMP with half the number of cell boards in the frame. Sun has hinted that there will be Jupiter boxes with four or eight Olympus sockets, and it stands to reason that machines with 16 or 32 sockets will also be available.
When it comes to performance, Mr McCormack says that he is advising customers that the dual-core Olympus processors will deliver about 2.1 times the performance running in the Jupiter servers as the single-core 1.35GHz Sparc64 V processors delivered in the current PrimePower XA frames. The Jupiter frames, by the way, will not be able to support the older Sparc64 V processors, since the move to dual-core chips coincided with an upgrade in the memory and bus architecture in the PrimePower line.
Further down the road, by about the middle of 2007 or so, Fujitsu will move the kickers to the Olympus chips to a 65 nanometer process and expand the on-chip caches and crank up the clock speed, too. It is also remotely possible that Fujitsu adds more cores as it continues to shrink the chip processes.
In the meantime, for those customers not ready to upgrade to the Olympus-Jupiter products, Fujitsu intends to crank up the clock speed a little on the existing Sparc64 V processors. Last June, Fujitsu boosted the clock speed on the single-core Sparc64 V chip to 1.89GHz and rolled that chip into its midrange and high-end PrimePower line between September and December. By the end of this year, Fujitsu has committed to boosting the Sparc64 V chip to 2.16GHz, and it has said in the past that it can increase cache size on the chip and crank the clock speed up to about 2.4GHz.
On the Itanium front, Mr McCormack also said that the so-called Mission Critical Intel Architecture server, the future Itanium-based machine that Fujitsu-Siemens has been working on for more than two years, is also on track for its own launch in the middle of this year. The server has not yet been named, and Mr McCormack quips that he is open to suggestions. (PrimeRib? PrimeDirective? PrimeNumber? PrimeSuspect? You decide.)
Like the PrimePowers, the future Itanium server will be based on a flat SMP architecture. It will scale from 4 to 32 sockets and from 64GB to 1TB of physical main memory. The system will have fully mirrored buses and plenty of autonomic operations features, such as improved fault protection and fault detection, that are typical of mainframe-style and high-end RISC/Unix computing and generally missing from the typical x86 entry or midrange servers.
The main memory in the systems will also be mirrored, and the machine, which will support Windows and Linux, will be able to have physical partitions. He says that Fujitsu-Siemens opted for Itanium – and not Xeon or Opteron – for its high-end Intel server because of the resiliency features Itanium has and that these other chips do not.
While Mr McCormack does not say this, it is clear from the timing of the launch of its Itanium server that Fujitsu-Siemens was counting on the dual-core Montecito Itanium processors to make a big splash. These processors were originally expected around mid-2005, but last fall were pushed out to initial deliveries at the end of 2005 and volume shipments in early 2006.
But because the existing single-core Madison Itanium 2 and dual-core Montecito processors are socket-compatible, the Fujitsu-Siemens Itanium box can launch this summer using the Madison chips. The Montecito chips will come in handy for virtual machine partitioning, too. The Montecitos support Intel’s Vanderpool VT hardware-assisted virtualization technology; the Madisons do not.