Special Processors to Drive IoTs and Wearables (Part 2 of 2)

V.P. Sampath is a senior member of IEEE and a member of Institution of Engineers India. He is working in FPGA Design house. He has published international papers on VLSI and networks

3498
Advertisement

In the world of ultra-low-power applications, power is everything. There are three basic types of systems depending on the sources of power:

1. Systems powered directly by energy harvesting (for example, a solar cell)

2. Systems powered by a battery, which, in turn, is charged via energy harvesting

3. Systems that just have a battery

It’s the size of energy harvesting and/or energy storage (battery) components that ultimately determines the form factor, size and weight of an ultra-low-power device. Peak power and energy consumption impact harvester and battery size calculations in different ways depending on the type of device.

Harvester with battery
Fig. 6: Harvester with battery

Current approaches for determining peak power and energy

There are three basic approaches for determining the peak power and energy requirements of an ultra-low-power processor for a given application. For a very conservative upper bound, look at the datasheets for the processor, which tell the peak power that can be consumed by the hardware. Run a stress mark—an application that attempts to activate the hardware in a way that maximises peak power or energy. A stress mark may be less conservative than a design specification, since it may not be possible for an application to exercise all parts of the hardware at once. Perform application profiling on the processor by measuring power consumption while running the target application on the hardware.

Things tend not to work out well if a processor tries to operate outside of the peak power and energy bounds available to it. So the best practice is to add a guard band (power buffer) to any profiling-based results. A typical guard banding factor might be 33 per cent. In other words, take the profiling results, add 33 per cent and provision for that.

Input-independent peak power and energy profiling

The core idea is to run a symbolic simulation of an application binary on the gate-level netlist of a processor. Most ultra-low-power systems tend to have simple processors and applications, making simulation feasible. For example, even the most complex benchmark may complete full simulation in two hours.

During simulation, a special X value is propagated for all signal values that can’t be constrained based on the application. Start with all gates and memory locations not explicitly loaded with the binary set to X. Any input values during simulation are also set to X.

As simulation progresses, the simulator dynamically constructs an execution tree describing all possible execution paths through the application. If an X symbol propagates to the inputs of the program counter (PC) during simulation, indicating an input-dependent control sequence, a branch is created in the execution tree. Normally, the simulator pushes the state corresponding to one execution path onto a stack for later analysis and continues down the other path.

Types of harvesting
Fig. 7: Types of harvesting

At the end of this process, you know the activity of each gate at each point in the execution tree. A gate that is not marked as toggled (0 to 1, or 1 to 0) at a particular location in the tree can never be toggled at that location in the application. You can use this knowledge encoded in the execution tree to generate peak power requirements as follows:

1. Concatenate all of the execution paths into a single execution trace.

2. Assign all of the Xs in the trace such that power for each cycle is maximised. Power is maximised when a gate toggles, but a transition requires two cycles—one to prepare and the other to make the transition. Since you don’t know the best way to align transitions with cycles, two separate value-change dump (VCD) files are produced; one maximises power in all even cycles and the other maximises power in odd cycles.

3. Combine the even and odd power traces into a single peak power trace by taking power values from even cycles in the even trace, and odd cycles in the odd trace.

4. Peak power requirement of the application is the maximum power cycle value found in the peak power trace.

5. When making this calculation, for input-dependent branches the algorithm always take the most expensive one. For loops where the maximum number of iterations can be determined, simply take the energy for one iteration and multiply it by that maximum number. If neither is possible, it may not be possible to compute the peak energy of the application; however, this is uncommon in embedded applications.

6. Results. By accounting for all possible inputs using symbolic simulation, this technique can bound peak power and energy for all possible application executions without guard banding. Peak power requirements reported by the technique are 15 per cent lower than guard-banded application-specific requirements, 26 per cent lower than guard-banded stress-mark-based requirements, and 27 per cent lower than design-specification-based requirements. On an average, the peak energy requirements reported by the technique are 17 per cent lower than guard-banded application-specific requirements, 26 per cent lower than guard-banded stress-mark-based requirements, and 47 per cent lower than design-specification-based requirements.

Advertisement


SHARE YOUR THOUGHTS & COMMENTS

Please enter your comment!
Please enter your name here