ARM's lead processor architect has told us that the company has thought about creating an inexact processor; a processor that curtails precision in the interest of saving power.
Richard Grisenthwaite, lead processor architect with ARM, has said the company has thought about creating an inexact processor -- a processor that curtails precision to save power. The technique has echoes of fuzzy logic and probabilistic processing and uses reduced accuracy of multiplication and addition but manages the probability of errors building up.
Grisenthwaite, vice president of technology and an ARM Fellow, has been responsible for the ARM architecture for the last ten years or so. He is associated with the ARMv7 instruction set architecture and the Cortex range of processor cores, with the move to 64-bit computing with ARMv8, and the "big-little" strategy now being deployed in both ARMv7 and ARMv8. Big-little pairs a performance-optimized processor core (big) with a power-optimized core (little) and the software load is allowed to transition between them as part of a dynamic voltage and frequency scaling regime.
I met Grisenthwaite recently at an ARM-organized analysts' conference and so took the opportunity to ask him about inexact computing and the possibility of its deployment by ARM.
Inexact processing is not new but seeks to address applications that do not require the full precision that is conventional in software but is therefore responsible for much of the power consumption without benefit. The technique is relevant in audio and graphical subsystems where the ear or eye often cannot perceive the full computed resolution and where human beings are good at error compensation. It is also applicable in some database and networking applications and control algorithms.
Different strokes for different apps
In the past it has been the case that because some applications required full precision -- and software was written using precision data types -- it was necessary that a computer's single processor be capable of full precision. But with heterogeneous multicore processing and hardware offload entering the mainstream, there is now an opportunity to provide different processor cores within a single system-chip optimized for different classes of application and providing the associated precision.
Earlier this year EE Times reported on a prototype inexact processor built by an academic team that demonstrated energy savings of 90 percent for particular applications where answers deviated from the correct value by 0.25 percent on average. (See: Inexact processor is more power efficient.)
So is there scope for ARM to produce a processor architecture or micro-architecture that somehow uses "pruned" ALUs and holding registers with fewer bits to achieve acceptable answers at much lower power consumption?
"ARM has thought about that. But it is quite hard to see how we would deploy it under our business model. It tends to be application area specific and we have to build cores that go to many applications and markets," Grisenthwaite told me. "And you would have to put it alongside a core designed to run conventional precision and legacy code."
My response was that this could be a good fit with ARM's computing philosophy of having multiple cores offering different required levels of application performance and energy efficiency and that an inexact processor could be run with other cores under a big-little or heterogeneous computing scheme. "It is something ARM has looked at but there's nothing to say about it" was Grisenthwaite's reply and clearly a concluding remark.
Does that mean ARM is working on an inexact processor? Your interpretation of Grisenthwaite's remarks is probably as good as mine -- but perhaps the company should be?