In an interesting twist, delegates at the International Solid State Circuits Conference (ISSCC) yesterday agreed that microprocessor speed is no longer a measure of performance.
Clock speed has been one of the most identifiable aspects of processor performance for a decade. This needs to change as process technologies shrink and alternative methods of increasing performance should be examined, experts agreed.
"In the past, performance has tracked directly with frequency. That's clearly ending," said Philip Emma, manager of systems technology and microarchitecture at IBM.
Building a high-performance microprocessor involves a series of tradeoffs, and there is no one accepted way to increase said performance. Some designers increase the chip's clock speed, some choose to improve the bandwidth of interconnects, and others explore ideas such as multithreading and multicore designs. Most chip companies use some combination of the three to deliver ever-increasing levels of performance.
With the chip industry in the midst of a process technology generation leap, concerns about power dissipation are making some chip designers wonder if the days of higher and higher clock-rates are done. Concerns about power dominated an earlier ISSCC presentation of cutting-edge processors, where just about every presenter addressed the issue of managing power dissipation.
"If you push frequency further, you have to spend power. If we push frequency any further, we're not going to be able to cool these things," said Alisa Scherer, a fellow with Advanced Micro Devices.
The lone cheerleader for increasing gigahertz was Doug Carmean, principal architect on the Pentium 4 project at Intel. Bringing higher and higher frequencies to its Pentium 4 chips has served Intel well for several years, and the company plans to bring the newest version of the Pentium 4 to 4GHz by the end of the year.
Increasing clock speed is still the best way to advance performance for certain applications on the desktop side of the world, such as three-dimensional rendering, Carmean said. He acknowledged that extracting performance from single-threaded workloads is difficult, but it works, and frequency is the most easily understood aspect of performance, he said.
Most PC consumers don't understand exactly what makes their machine perform, but can be trained to recognize other aspects of chip performance, Scherer said, alluding to the various aspects of automobile performance that enter into a purchase decision.
"For every dollar you spend trying to manage power, you could spend on marketing your overall performance," Scherer said.
For many applications in the server world, frequency is not the answer, said Marc Tremblay, vice president and fellow with Sun Microsystems. Sun just released its first dual-core processor, and is working on multiple-core designs where each core runs several threads, he said.
By moving to multiple cores, chip designers can also reduce the footprint of the chip die while maintaining or improving performance, said Hisashige Ando, chief technology officer with Fujitsu's enterprise systems group. Multiple cores each running at reduced clock speeds can also save power, he said.
Power concerns loom over every current processor design team, and the situation will not get better any time soon, said Mark Horowitz, professor of electrical engineering and computer science at Stanford University. Designers will have to learn to live within a power budget while still charged with increasing performance, he said.