"Megahertz was how we increased performance in the beginning, but we've been saying for a couple years now that that will run into some fundamental problems," said Frank Spindler, a vice president of Intel's corporate technology group who oversees IDF. "If you keep scaling frequency, you begin to see diminishing returns in processor performance while you see increasing issues in power consumption." Spindler added.Uh huh. For years now.
In fact, Intel just recently found itself facing a two-pronged dilemma. On the one hand, semiconductor scaling had apparently hit the rocks at 90nm, throwing Moore's law into question. On the other, CPU engineers had run out of obvious tricks for achieving increased linear performance by dedicating more transistors to processing logicand especially of doing so without creating darn near unmanageable complexity. The solution, Intel declared in a rare and public major rewrite of its roadmap, was more parallelism through placing multiple CPU cores on a single chip.
What we've been wondering, in part, is which of Intel's processors would become dual-core parts, and given the obvious problems with the Prescott Pentium 4's size and performance, which technology would become the foundation for Intel's next-gen desktop parts. The Pentium M, derived from the Pentium III but compatible with the Pentium 4 bus, has been one leading candidate for both jobs. A Pentium M-derived dual-core chip would quickly give Intel a credible alternative to AMD's Opteron, with a similar number of pipeline stages, a much smaller die than Prescott, and built-in technology for managing power and heat. However, rumors persisted that Intel was prepping a dual-core processor based on the Pentium 4's Netburst architecture, and we headed into the Intel Developer Forum this week expecting to see a live demo of some Intel processor running with dual cores.
Turns out that Intel has gone crazy for dual cores across the board. The first chip it demoed this week was a dual-core Itanium called Montecito. With 1.72 billion transistors and 27MB of total onboard cache, this monster requires a gargantuan heatsink and thermoelectric cooler just to keep from shooting hot magma out of the CPU socket. (Itanium is obviously becoming an ever-smaller-volume niche product over time, especially with the advent of 64-bit x86 alternatives.) Then Intel demoed a dual-core desktop processor based on the Netburst architecture and running with a 915 chipset, proving it has succeeded in gluing together two Prescotts. This demo, coupled with word that Intel plans to bring its SpeedStep power saving/clock throttling technology to desktop chips, suggests to me that a dual-core chip based on the Netburst microarchitecture will be Intel's future desktop CPU.
The Pentium M won't be left out of the double-headed action, though. Intel is also prepping a chip code-named Yonah, a dual-core version of the Pentium M, that can turn its second core on and off as needed to conserve power.
Observing from afar, the news coming out of this week's IDF suggests several things about the shape of Intel's newfound religion. First, the desire to move to more parallelism is apparently quite heartfelt. All three of Intel's major CPU architectures are making the move at once. However, this new direction isn't accompanied by other big shifts in Intel's approach to the world. The company is apparently quite happy with offering three very different CPU architectures to three different markets, and all three of its current architectures appear to have a future in the dual-core world. The Pentium M won't bleed over into the desktop world, at least not yet, and the Itanium will just keep moving upmarket to capture the big bucks in the enterprise world.
Most notably, it seems Intel's faith is Moore's Law is largely unshaken. AnandTech has an interesting article detailing Intel's plans to move to new process technologies, and all of Intel's plans seem to be riding on continued success in this area. The relatively large Pentium 4 and Itanium architectures remain in favor, and they are scheduled to grow to ever-larger transistor counts. As the article points out, the Pentium 4 has moved from 256K of L2 cache to 512K and then 1MB, and a 2MB L2 cache is already planned for early 2005. Larger caches and increasing parallelism have replaced rising clock frequencies as the primary means of exploiting Moore's Law, but the old dictum about transistor counts doubling every 18 months remains Intel's guiding light.
So we won't likely see a radical transformation of Intel's CPU architecture plans, at least in the near term, beyond the efforts now underway to staple together pairs of CPU cores on a single die and sell them to their intended markets. That's what I take from the reports coming out of IDF this week, at least, and it does offer some clarity about exactly how Intel's conversion to its newfound religion will play out.