Attack Of The 64-Chip Mainframe
If you look at the surface of your coffee as it sits quietly in its cup on the desk, you may spot some tiny, rhythmic ripples. In Jurassic Park, that's the sign of an incoming Tyrannosaurus Rex: in enterprise IT, it's the seismic early warning of the return of the mainframe.
Or so IBM would have us think. Today, it announced its z10, a 64-chip quad-core giant that can replace "1500 x86" servers (IBM quotes "On Line Transaction Processing Relative Processing Estimates (OLTP-RPEs): Derivation of 760 Sun X2100 2.8 Opteron processor cores with average OLTP-RPEs per Ideas International of 3,845 RPEs and available utilisation of 10% and 20 RPEs equating to 1 MIPS compared to 26 z10 EC IFLs and an average utilisation of 90%." in support of this claim. And who are we, etc...). Big energy, license and operational savings are promised.
IBM is being extremely bullish about mainframes. Word in the high-performance computation jungle is that packs of highly-trained sales velociraptors are stalking any large, lumbering organisation with a large, lumbering budget and missing no chance to sink serrated teeth into unguarded flank. There's something almost majestic about the sight of a company as prehistoric as Big Blue aggressively reclaiming a niche long thought of as endangered.
Other predators are available. In particular, over at Japanese web site PC Watch you can see quite a few details of Intel's big server chip architectural upgrade, Nehalem. (Don't worry about not reading Japanese - there are enough acronyms and interesting pictures to carry the story). Nehalem is a brand-new x86 core, and will live in a number of chips - the most interesting of which is code-named Beckton (*).
According to PC Watch, Beckton has eight cores, lashings of cache and most importantly Quick Connect - Intel's own next-generation interconnect. That leads to all manner of interesting configurations, including -- again according to PC Watch -- one eight-chip, sixty-four core design -- which reminds me of nothing so much as the Jewish mystical Kabala Tree of Life, though I don't think Haifa's involved. Four of those servers would hit the z10's core count - and, of course, there's every chance four eight-core Nehalems will be replaced with one 32-core chip later in the cycle.
As IBM will tell you at great length, high-performance computing is just as much about moving data around as it is doing the raw computations, which is why you get such benefits from having your CPUs tightly coupled in a box instead of stringing together hundreds of separate servers over a network. AMD has also done well out of its own Hypertransport interconnect during the time it's taken Intel to get around to the idea.
But Intel is coming. Expect a great deal more complex benchmarking, especially next year when - as Intel hopes - multiprocessor Nehalem systems start to get within shelling range of machinery like the z10. But IBM has at least 18 months lead and is a hell of a competitor, and there's lots that can go wrong with ultra-fast interconnects: that's where the most firepower will be directed in the coming war.
(*) The Beckton after which the chip is named is probably a small town in Wyoming. It is probably not the scuzzy post-industrial wasteland in East London best known for its sewage works, through which most of the poo in London passes on its way to annoy the French.