Last week there was an inaugural ARM Research Summit.
https://developer.arm.com/research/summit
I gave a talk on Power & Compute Codesign for “Little Digital” Electronics.
Here are the slides of this talk:
https://www.staff.ncl.ac.uk/alex.yakovlev/home.formal/Power-and-Compute-Talk
Here is the abstract of my talk:
Power and Compute Codesign for “Little Digital” Electronics
Alex Yakovlev, Newcastle University
The discipline of electronics and computing system design has traditionally separated power management (regulation, delivery, distribution) from data-processing (computation, storage, communication, user interface). Power control has always been a prerogative of power engineers who designed power supplies for loads that were typically defined in a relatively crude way.
In this talk, we take a different stance and address upcoming electronics systems (e.g. Internet of Things nodes) more holistically. Such systems are miniaturised to the level that both power management and data-processing are virtually inseparable in terms of their functionality and resources, and the latter are getting scarce. Increasingly, both elements share the same die, and the control of power supply, or what we call here a “little digital” organ, also shares the same silicon fabric as the power supply. At present, there are no systematic methods or tools for designing “little digital” that could ensure that it performs its duties correctly and efficiently. The talk will explore the main issues involved in formulating the problem of and automating the design of little digital circuits, such as models of control circuits and the controlled plants, definition and description of control laws and optimisation criteria, characterisation of correctness and efficiency, and applications such as biomedical implants, IoT ‘things’ and WSN nodes.
Our particular focus in this talk will be on power-data convergence and ways of designing energy-modulated systems [1]. In such systems, the incoming flow of energy will largely determine the levels of switching activity, including data processing – this is fundamentally different from the conventional forms where the energy aspect simply acts as a cost function for optimal design or run-time performance.
We will soon be asking ourselves questions like these: For a given silicon area and given data processing functions, what is the best way to allocate silicon to power and computational elements? More specifically, for a given energy supply rate and given computation demands, which of the following system designs would be better? One that involves a capacitor network for storing energy, and investing energy into charging and discharging flying capacitors through computational electronics which would be able to sustain high fluctuations of the Vcc (e.g. built using self-timed circuit). The other one that involves a switched capacitor converter to supply power as a reasonably stable Vcc (could be a set of levels). In this latter case, it would be necessary also to invest some energy into powering control for the voltage regulator. In order to decide between these two organisations, one would need to carefully model both designs and characterise them in terms of energy utilisation and delivery of performance for the given computation demands. At present, there are no good ways for co-optimising power and computational electronics.
Research in this direction is in its infancy and this is only a tip of the iceberg. This talk will shed some light on how we are approaching the problem of power-data co-design at Newcastle, in a series of research projects producing novel types of sensors, ADCs, asynchronous controllers for power regulation, and software tools for designing “little digital” electronics.
[1] A. Yakovlev. Energy modulated computing. Proceedings of DATE, 2011, Grenoble, doi: 10.1109/DATE.2011.5763216