Beyond the Hype

By Jerry Pournelle

May 23, 2005

(Beyond the Hype :  Page 1 of 1 )

Column 298 (Continued from the Previous Week)

HyperTransport, Hyper-Threading, and Hype

There is multi-processing in your future. Depend upon it.

Moore's Law states, roughly, that system capabilities double about every eighteen months. This is an empirical rule that has been observed since the 1970s, and predictions based on this "law" have proved out. The original formulation was based on a more underlying empirical observation, that the number of transistors that could be put in a very large scale integrated chip doubled on the same time scale. Note that the two "laws" are not quite the same, and that while the second statement implies the first, it's not a certain cause and effect.

The remarkable thing has been how well this prediction has held up, particularly recently when attempts to improve CPU performance through cranking up the clock speed became counter-productive: Above 3 GHz increasing clock speed produced far more heat than computational power. The clock speed game ended with Prescott, which was supposed to go up to 5 GHz, but which will probably never top 4.0 GHz.

If speeding up one CPU can't do the job, how can designers take advantage of increased transistor densities to get higher performance? The obvious answer is through multiple processors. Instead of more speed, you break the problem up into parts and have a number of slower machines each do a part of the job at the same time, then integrate the results.

At the extremes required for scientific computing, this approach has been known as "massive parallel processing," and showed up in our world of small computers as "Symmetrical MultiProcessing" or SMP. Most of us experienced it, if at all, as "dual processing." Moving to four processors was tougher. Using massively parallel systems for general purpose computing proved to be really hard, as was stated succinctly by John McCarthy: "Such systems tend to be immune to programming." Over time many of the problems were solved—particularly in graphics processing—and Intel has featured multi-threaded processing, called Hyper-Threading, for some years now.

 Page 1 of 1