Intel adapts to changes ahead

SAN FRANCISCO--If processors are going to continue to increase in performance, they'll have to be a lot more flexible in the future.

With that in mind, Intel is seriously considering adopting techniques from the reconfigurable processing world for future products, Pat Gelsinger, chief technology officer at the company, said at the Intel Developer Forum here. The company is hoping the techniques will help it to better address ways to boost chip performance without drastically increasing power consumption.

Reconfigurable processors adapt to their workloads. Different subsections and pipelines of a chip, for instance, are activated to render a large graphical image or to run a database. Chipmaker Xilinx has been a leading proponent of this approach. Adaptive processing also is at the heart of Sun Microsystems' Throughput Computing initiative.

With adaptability, a chip can "go fast, do less speculation (a design convention in which a chip tries to predict future tasks that can consume excess energy) and run some cores more slowly," Gelsinger said.

Intel has already posted research results using adaptive techniques in networking processors that show that certain tasks can be accomplished far faster because less intermediate steps are needed.

"There will be a lot of borrowing of ideas" from papers and companies that have examined adaptability before, Gelsinger said.

Adaptability is part of an overhaul of microprocessor designs that began at the turn of the decade. For years, chip designers raised processor performance by increasing clock speeds (measured in gigahertz) and adding transistors. Those techniques, however, have expanded the amount of energy required to run processors. Roughly 20 percent to 40 percent of the energy consumed by modern 90-nanometer chips gets lost through leakage, Gelsinger noted, an assertion backed up by research reports from companies.

Not only is it expensive, excess energy consumption increases the amount of heat processors exude. Heat can damage internal computer parts, scramble signals, and force computer designers to come up with novel and expensive parts for removing heat from the inside of boxes.

As a result, Intel and other microprocessor makers are going to have to get performance gains from techniques such as multicore chips and threading.

"Power, memory latency and (clocking) delays are going to curb the rate of frequency growth," Gelsinger said. "There needs to be a major architectural paradigm shift."

In the meantime, of course, processor speeds will continue to accelerate. Intel's Prescott will hit 4GHz by the end of the year.

Invite Michael Kanellos into your in-box Senior department editor Michael Kanellos scrutinizes the hardware industry in a regular Enterprise Hardware column that ranges from chips to servers and other critical business systems.

Do consumers and businesses still need higher performance? Of course, Gelsinger said. Historically, analysts and others have always questioned the need to speed up chips. But new applications always come along that tax hardware.

The future will be no different. Search engines that can find images will be used by doctors to compare a patient's MRI scan or X-ray with clinical examples, he said.

Graphics also will change. Now, graphics chips render a scene by breaking it down into microscopic triangles. While that can result in realistic graphics, these images will always be approximations.

In the future, graphics specialist will use a more accurate technique called ray tracing, which involves simulating how light bounces of surfaces. Ray tracing has actually been around for years, but the computational power to run tracing applications has been astronomical.

inTrace, a start-up in Germany, is already selling ray tracing software to Volkswagen and Airbus and is talking to Boeing, inTrace founder Phillipp Slusallek said at the forum.