PORTLAND, Ore. By solving the parallel processing problem with application-specific languages, Intel Corp. Senior Fellow Justin Rattner predicted that future multicore processors will divide and conquer rather than depend faster single-processor speeds.
Speaking at the Intel Developer Forum Thursday (March 3), Rattner focused on how Intel is addressing parallel processing and bandwidth challenges. Intel reported successful fabrication of its first samples of a
dual-core Pentium, called the Extreme Edition, and plans to begin delivering it during the first half of 2005.
Intel also said it has ten additional multicore processors in the pipeline for 2006 and beyond that will address different segments of the server, network, desktop and
mobile markets. Rattner's job is to make sure the parallel processing capability of the multicore engines is fully utilized by software.
"The general-purpose approach we had been pursuing hadn't made much progress despite many years of effort. But now I think we have a winning approach. That
is, we want to solve parallel processing with application-specific languages, as opposed to trying to create a general-purpose, parallel programming language."
Rattner demonstrated a compiler written in a high-level language called "Baker" that incorporates domain-specific
features to manage parallel packet processing on a network. Previous parallel processing languages had hardware-specific features, but the new application-specific approach, according to Rattner, enables the
compiler to produce efficient binaries that divide up specific parallel processing problems.
Rattner said the single-processor approach would have run out of steam before clock speeds could be ratcheted up ten-fold over current speeds. By then, Intel will have its parallel processing software handling terascale applications by running them on chips with from 16 to 256 cores on a single die. The multicore chips
will be cheaper to make and consume much less power, but new software must harness parallelism to smooth the transition to future terascale applications.