ARM unveils revolutionary new microarchitecture
Cambridge UK superchip architect ARM has unveiled technology that is set to expand the possibilities for Artificial Intelligence (AI).
ARM says DynamIQ is probably the biggest micro-architectural shift since the business announced 64-bit ARMv8-A in 2011. It represents entirely new grounds for multicore microarchitecture that will serve as foundation for all new Cortex-A processors – both big and little.
ARM Cortex-A multi-core processing is optimised for automotive, networking and server capabilities with reliability and safety. And it is super-fast, says the company.
ARM expects that its partners will ship 100 billion chips in the next five years. To put the significance of that number into perspective, over the past four years ARM partners shipped 50 billion chips.
Of those next 100 billion, Cortex-A figures to be an even more prominent factor due to this new technology which opens opportunities to go far beyond mobile and into other applications.
ARM says new dedicated instructions for AI and ML will deliver a 50x boost in AI performance over the next three to five years. ARM is saying this is a conservative number as its only building out projections based on AI algorithms they know about or have access to. They say there is much more to come.
In terms of the server/cloud computing environment, DynamIQ will be the second ARMv8-A ISA enhancement ARM has announced to-date; the other was the scalable vector extensions (SVE) announced at Hot Chips.
The fact ARM can now scale to eight cores on a single cluster is huge for networking environments. ARM now scales higher with significantly reduced latency through more efficient CPU to CPU communications.
In automotive, the technology promises more safety features for safer autonomous cars and industrial systems. The AI instructions will also help on this front.
Given the priorities around heterogeneous computing and device-based AI on mobile devices as well as notebooks, this architecture is much more flexible and configurable than ARM’s previous iterations of big.LITTLE.