Slashdot Mirror
DARPA's Latest Chip Is Designed To Be Bad At Arithmetic (technologyreview.com)
Reader holy_calamity writes: Pentagon research agency DARPA has funded the creation of a chip incapable of correct arithmetic, in the hope of making computers better at understanding the real world. A chip that can't guarantee that every calculation is perfect can still get good results on many problems but needs fewer circuits and burns less energy, says Joseph Bates, cofounder and CEO of Singular Computing. The S1 chip can process noisy data like video very efficiently because it doesn't need the extra circuits or operations needed to ensure every mathematical operation is performed perfectly. This summer DARPA will put five prototype computers, each equipped with 16 of the inexact S1 chips, online for researchers to experiment with.
Yes, I guess you can make arithmetics slightly faster when you allow errors, but is that where todays CPUs spend a lot of time?
Not so much in error checking, but in the choice of the algorithm itself.
As an example, Quake 3 famously used a crazy-fast inverse square root routine. It didn't give an exact answer, but rather, one "close enough" to suit its intended purpose (calculating surface normals for reflections) in software, in a quarter of the time FPUs of the era could get an answer using dedicated hardware. The FPU would always give a much more accurate answer, but not every use needs a much more accurate answer.