As IBM announces the POWER7 today, Sun-Oracle will be telling their fleeing customer base that IBM finally, albeit after five years, validates Sun’s Chip Multi Threading (CMT) processor architecture.
Sun-Oracle will attempt to equate their five year old, 8-core 4 threads per core CMT processor called the UltraSPARC T1 (Niagara-1) with today’s IBM POWER7. Indeed, the IBM POWER7 has 8 cores and 4 threads per core - but that is where the numerical similarities end and Sun generated FUD (Fear, Uncertainly, and Doubt) begins.
Sun-Oracle will argue that not only does it market a second-generation CMT processor, the UltraSPARC T2 (Niagara-II), with 8 cores and 8 threads per core but that its next incarnation will have 16 cores with similar threading. If one were to set as equivalent the number of cores on a piece of silicon as the test of greatness, one might note, among the multitudes, Intel’s 16-core, multi-threaded network processors, IXP2400/2800 available well before Sun’s CMT, or Octeon’s encryption processor with 16 MIPS64 cores. Alas, Sun would say that both these processors, and many others like them, are application specific.
However, Sun-Oracle’s current CMT processors are also application specific. They only perform well in thread-rich environments. Such environments are typically web servers and lightweight databases where strong single thread performance is not necessary. One need only note the types of benchmarks Sun publishes - and those it does not - for confirmation of Sun’s CMT application specific resonance.
Sun attempted to design and Texas Instruments manufacture a CMT processor that would address both thread-rich and heavy single threaded execution requirements. Internally it was called the ROCK processor. That project ended in failure. Its chief architect left Sun and joined Microsoft last year. The reason Sun attempted to design such a CMT processor is because many of today’s applications still require swift execution of heavy single threads. Sun’s available CMT processors are so poor at executing single threaded code that it doesn’t even publish industry-standard single core benchmarks for their processors. POWER7’s published industry-standard benchmarks speak for themselves.
Both IBM and Intel could have easily designed, manufactured, and marketed processors that were both highly multi-cored and multi-threaded but would have done so by sacrificing the execution quality on a vast array of standard single-threaded data center applications. In contrast, it was not necessary for IBM to sacrifice the execution quality of existing data center applications. IBM evolved its multi-core RISC architecture beginning with its dual core POWER4 in 2001 to today’s 8-core POWER7 with a continual positive impact on data center execution quality and price-performance.
There is no better indication on how divergent multi-core and multi-threaded processor architecture and performance can be than to note that DARPA selected IBM’s POWER7 for its Supercomputing Grand Challenge (see: http://www-03.ibm.com/press/us/en/pressrelease/20671.wss). Sun was dropped from the competition - based on the broken promise of the ROCK processor by 2010. (See: http://m.channelregister.co.uk/2006/11/21/darpa_petascale/). Today’s announced POWER7 is part of DARPA’s Petascale Challenge, not Sun’s non-existent ROCK and certainly not its little brother, Sun’s UltraSPARC T2 processor.
David Davidian 270001BV1W firstname.lastname@example.org Tags:  darpa petascale cmt sparc sun-oracle power7 4,491 Visits