Guild Companies

Japanese server giant Fujitsu, which, with its alliance partner Siemens, has big aspirations in the Unix, Windows, and Linux enterprise server businesses, last week announced that it would be working with Intel to develop large Wintel and Lintel servers that will compete against Unix servers, including Fujitsu's and Siemens' own PrimePower Sparc-compatible machines. The future Fujitsu servers will be based on 32-bit Xeon and 64-bit Itanium processors and will be based in part on technology in the PrimePowers, which scale to 128 processors.

According to sources at Fujitsu and Intel, the development plans these two companies have put together call for a big box based on the follow-on to the current "Gallatin" Pentium 4 Xeon MP processors, on which Intel is just starting to ramp up production. (Oddly enough, no one seems to know the code name of this future Xeon MP chip yet.) It is hard to say how big the Xeon-based machine will get, especially since neither Intel nor Fujitsu is committing to anything specific at this point. Richard Dracott, group director of enterprise computing at Intel, would say, however, that the Xeon-based server would ship at the end of 2004; when pressed about whether it would scale to 128 processors in a single system image, Dracott said that was not likely. But scaling to 32 processors is likely, and a 64-processor machine is not out of the question. If the Gallatin chip eventually has a kicker that packs dual cores on a single chip, the Fujitsu-Intel designs could scale to 128 processors. This seems the only likely way the 128 processors in a single image is possible using Xeons over the next few years, just because of the limits to scaling beyond 64 processors in most server architectures. Adding processors beyond that would be very inefficient, and only useful for partitioned machines that are used for server consolidation.

Perhaps more significant for Fujitsu and Siemens, sometime in 2005 Intel and Fujitsu will have completed enough development work so that Fujitsu and Siemens can ship a 128-way machine based on the dual-core "Montecito" Itanium processor that was added to the Intel chip roadmap two weeks ago . We guess that Intel is targeting 2.5 GHz to 3 GHz with the Montecito Itaniums, which would give each core about 30 to 40 percent more oomph than the Madison-II Itanium processors that are expected in 2004 and which we expect to hit around 2 GHz, if Intel can swing it. Suffice it to say that 128 of these 2.5 GHz or so Montecito processors will pack a serious wallop. Without any of the architecture details, it is hard to say how much, but it will be much larger than any Wintel or Lintel server on the market today, and similarly much more powerful than any current RISC/Unix server that is out today or due in 2003 or 2004.

The future Xeon and Itanium servers from Fujitsu and Siemens will be sold by Fujitsu in Japan, by the Fujitsu Siemens Computers alliance in Europe, and by Fujitsu Technology Solutions in North America. The machines will run a seriously hacked version of the open-source Linux environment, with Fujitsu and Intel engineers putting in lots of time extending the current Linux kernel to scale to 128 processors. (It tops out at about eight processors in the Linux 2.4 kernel today.) While Fujitsu and Intel didn't want to talk about this, they did admit that the machines would also run a future version of the Windows operating system, which means the "Whistler" version of Windows, now called Windows Server 2003 and due to ship at the end of April, on the Xeon box that is coming out in 2004. Windows support on the 128-way Itanium machine in 2005 could end up being the "Blackcomb" kicker to Windows Server 2003, or it might be an enhanced version of Windows Server 2003.

Fujitsu and Siemens will sell this future Wintel and Lintel product line alongside the current Primergy Intel-based servers they already have in their portfolio, and the two have no intention of ceasing the sales of their PrimePower clone Sparc servers, which run the Sun Microsystems Solaris operating system, or of stopping development of the Fujitsu Sparc64-GP processors that are at the heart of the PrimePower machines. The development work being done by Intel and Fujitsu mirrors that being done by Intel and Hewlett-Packard to create big Wintel and Lintel iron, and it is a matter of preservation for Fujitsu and Siemens to jump in this market. By getting Intel on board, Fujitsu and Siemens can get some of the best talent when it comes to dealing with the Intel architectures and also share some of the burden and cost of that development. The financial details of the relationship established by Fujitsu and Intel were not disclosed, and neither were the amount of people and intellectual property that the two would pump into the project.

Finally, neither company is clear about what will happen to the enhanced Linux that the two create for these future Fujitsu-Siemens machines. The odds favor some of this intellectual property remaining with Fujitsu and some of it being made open-source to appease the Linux community. Helping open Linux extend to 16-way or maybe 32-way SMP scaling, while keeping the intellectual property that allows 64-way to 128-way scaling, might be one way of dicing and slicing it. Another way might be to put the raw scaling capabilities into the open-source version of Linux but then providing the tuned versions of Linux, compilers, databases, and other middleware that allows the whole shebang to run much more efficiently only as a closed-source solution. This is what supercomputer maker SGI just did with its 64-way Linux server for HPC applications. Red Hat runs on the box right out of the box, but if you want it to run well you have to pay SGI for tweaked code.