Fast Systems Go with What Programmers Know
The Processor Forum used to be my favorite conference every year, because I was sure that the ingenuity of processor developers was defining the instruction set environments in which wed spend our time crafting code in the years to come. A briefing on developments in the Alpha or the Itanium or the Power processor family was a preview of an adventure in the offingan orientation to the language wed need to know as we made our way in a fascinating new place.
At some point, though, it grew obvious that the most promising destinations were all beginning to be places where they spoke the familiar language of x86. It was like visiting a foreign country, and discovering that the most educated residents were more interested in practicing their English on you than they were in listening to your attempts to speak their language. You could have a much more stimulating conversation in the language you both knew best, and x86 plays that role for an ever larger fraction of the developers who need to speak hardwares language at all.
Thats the message I take away from the Nov. 13 release of the latest Top500 list of supercomputing installations, scheduled for official presentation on Nov. 14 at the SC06 international conference of high performance computing and networking in Tampa, Fla. This list update marks an important shift, with IBMs Power-family processors falling behind AMDs Opterons.
If we group Intels Pentium and Xeon with AMDs Opteron and call them all x86 processors, we find them now representing the brains of 341 of the Top 500 high-performance computing installationsmore than two-thirds of those sites, and just over 50 percent of their aggregate computing power as measured by floating-point operations per second.
Power-family processors represent a little more than a third of the Top 500s aggregate capability; Itanium CPUs less than a tenth; the remaining 5 percent is divided among PA-RISC, SPARC, Alpha, NEC ... oh, yeah, and Cray. Remember Cray? The name once synonymous with the most wicked-fast computers on the planet? Now reduced to not much more than a footnote of rounding error on the worlds list of hot machines?
As Stanford University professor John Hennessy observed seven years ago, the industrys asset base of coder competence turns over much more slowly than its base of hardware. The continued expansion of the envelope of x86 performance, largely unforeseen a decade ago, gives us all the chance to have the interesting conversations that we need to have about threading and clustering and manageability and fault toleranceinstead of learning to say that the pen of my aunt is on the bureau of my uncle in a new instruction set every five years.
Tell me what conversations youd like to have in tomorrows code at firstname.lastname@example.org.
Check out eWEEK.coms for the latest news, reviews and analysis in programming environments and developer tools.