Ever Since the 6th Generation Core CPUs, Intel has aimed to swiftly and efficiently manage CPU power states. Knowing when to shift the CPU to its highest performance mode, how long to sustain it, and when to return to a low power state is crucial. This power management grows more nuanced with each Intel Core CPU iteration.
At the recent Hot Chips conference, Intel revealed an AI-powered approach. This method, based on internal testing, could lead to substantial power savings. Currently, the Dynamic Voltage and Frequency Scaling (DVFS) mechanism involves choosing from various CPU performance levels. The CPU exposes these to the operating system, which then selects a level based on the workload. The CPU then autonomously adjusts, translating parameters into hardware-managed P-states that determine frequency and voltage.
With Intel’s next-gen energy-efficient CPUs, Intel uses AI for significantly more nuanced power management. Intel trains the AI offline and can thus adjust the performance of the CPU significantly better than before.
Power management, however, exceeds swift CPU return to low power (RTH or Race To Halt); it encompasses rapid ramp-up to the required performance. Intel claims AI enables the CPU to reach the desired high-power state 35% faster than previous methods.
While the AI can’t learn from specific end-user behaviour, Intel’s trained AI marks significant progress, and it offers a more responsive yet energy-efficient UX. This optimises the CPU by delivering the required performance promptly and only when needed.
The initial desktop CPUs with this AI-trained power management are likely Intel’s upcoming Meteor Lake products, anticipated in 2024.