Chips that do one thing well have been hyped as a pathway beyond the end of the Moore's Law era. But the promise of these specialized microprocessors just suffered a reality check from Princeton University researchers. Led by David Wentzlaff, associate professor of electrical engineering, the team showed that, like with Moore's Law, the increase in usefulness of these so-called accelerators will reach a definite limit. And soon.
Fifth-year graduate student Adi Fuchs will base parts of his dissertation on these results. "Through the metric we developed, called the chip specialization return, we factored out the progress of the hardware over time and found that the gain from these specialized chips still greatly depends on transistor improvements," Fuchs said. "After transistors stop improving, chip specialization will run out of low- and medium-hanging fruit and end up delivering diminishing returns."
A recent report by IEEE Spectrum outlined this work by Fuchs and Wentzlaff, as well as their upcoming presentation at the 25th IEEE International Symposium on High Performance Computer Architecture. "Progress on accelerators... will hit a wall just like shrinking transistors will, and it will happen sooner than expected," according to the report.