In the previous blogs on this topic, we’ve seen that utilizing near-threshold voltage (NTV) saves incredible amounts of energy, theoretically up to 10x and in practice from 2x to 4x. But there is a price which makes some applications more suited for NTV than others. This is due to the inevitable performance (speed) loss of NTV as transistor current decreases with respect to operating voltage. While some applications require full speed all of the time, almost all IoT applications have widely varying performance requirements. Here, I’ll dig into one of the hottest IoT applications which happens to be an excellent fit for NTV: wireless audio hearables, also known as true wireless systems.
To be state-of-the-art in the very competitive hearables space, the system must include Keyword Spotting (KWS) such as Alexa or Siri. These are always-on systems: because the keyword can come at any random time, the system cannot be (completely) shut down. This already rules out long sleep times, the most common energy saving method of IoT systems. A typical KWS system consists of a feature extractor and a neural-network-based classifier, a form of artificial intelligence (AI). For energy efficiency, these are usually preceded by an energy/voice activity detector. This allows for the system to run the low-performance energy detection as the only always-on component and only wake up (via interrupt etc.) the main processor when energy resembling speech is detected. Of these, NTV is an excellent choice for the energy detector.
In a conventional energy-optimized system, the energy detector is often a hardwired block. Ultimately, time-to-market demands programmability as algorithms and architectures change, and anything hardwired severely limits this. One option would be a “big-little” type of system: a small CPU sharing memory and periphery with a bigger CPU such as an Arm M0 and an Arm M33, or two RISC-V cores. But even this solution has task-switching limitations on memory and switching time. If your software team gets to decide, all tasks will be run on the same core. Then there’s the extra silicon and verification costs that go into a multi-core solution.
Minima’s approach to a NTV system makes a single-core solution possible, one that can scale its energy together with its performance. As seen in Figure 1, using all of the energy curve (and not just a small sliver at the top) allows for optimizing energy no matter how much performance spread your application requires such as with keyword detection in hearables.
Figure 1: Minima’s approach to NTV operation enables a CPU to scale its energy (shown on the right) for simple parts of an algorithm as well as the more complex parts such as in hearables IoT applications.
Even better, Minima’s approach to NTV system maximizes the use cases of the CPU. Modern KWS algorithms are heavily optimized for small, embedded-class CPUs but today’s deep-submicron processes mean that often there is still room left at the top for you to design the system for more performance. So, when your algorithm and SW guys want more performance for a product with a bigger battery, you can reuse the system. For example, adding a 0.9V operating point in Figure 1 might allow for the same chip to be used as a speaker driver feedback DSP. Greater task granularity in your application may also be possible; for example, running different Bluetooth layers or neural-network layers at different operating points.
These examples of being energy frugal apply to a large number of other applications. Anywhere you need AI, there are probably energy-saving possibilities by using simpler algorithms part of the time enabled by Minima’s approach to a NTV system.Share this post via: