The invention discloses a method for metering power

consumption of a digital system-on-chip (SoC) for predicting battery life of a battery used by the SoC.

The invention also discloses an apparatus for performing the method of metering power consumption of a SoC.

Battery-powered devices are ubiquitous and cover a wide range of applications - from laptop computers to smart phones to small distributed sensors. Predicting end of battery life, or requirement to re-charge is important for most of these applications, either to provide a good user experience or prevent disruption of service. Especially small IoT devices pose a set of specific challenges, depending on their field of application: They are typically using small, non-rechargeable batteries, they are supposed to last for years in remote places, they might not be easily serviceable (e. g. buried sensors) and the application may still require an advanced indication of imminent battery loss .

Figure 1 shows a state-of-the-art battery monitoring 1, which is used in a variety of mobile devices, e.g. smart phones or laptop computers. The battery monitoring 1 comprises the following main parts: The actual battery Bat 2, which is typically a Lithium-Ion or Lithium-Polymer cell in today's devices. It further comprises the charge

electronic CE 3, which ensures a safe recharge of the battery 2 as well as a small series resistor R _meas 4 (in the order of lmW) to convert the current into a voltage. A differential amplifier A 5 is used to pre-process the voltage for the following units and an analog-to-digital converter (ADC) 6 converts the amplified voltage into a numerical value. A numerical integration (NI) block 7 accumulates the samples to calculate the drained charge from the battery, whereas a digital interface (DI) 8, typically I2C or SPI, uses the actual functional electronics to query the battery state. Last but not least, the actual functional circuitry 9 which is driven from the battery 2 can be anything from a smart phone's SoC to the motherboard of a laptop computer and it uses the digital interface 8 to learn about the battery state (and most likely communicates it to the user) using a built-in micro-controller.

As indicated in figure 1, the battery monitoring electronics 1 is often part of the battery pack itself and would thus be replaced along with it. Today, the charge electronics CE 3, the amplifier A 5, the ADC 6 with the numerical integration NI 7 and the digital interface DI 8 are all integrated into a single mixed-signal chip. It has the advantage, that this method of current metering covers all possible contributors of a platform and, since the metering chip is part of the battery pack 1, it can report the remaining battery charge right away, since it knows the battery size. But the

measurement resistor R _meas 4 is mostly a discrete part because of the high currents flowing through and the accuracy required for exact measurements. This approach, however, adds both cost and printed circuit board (PCB) area to the platform that would not be acceptable for small, low-cost IoT systems. It is therefore the object of the invention to find a possibility to observe and meter the battery life of small, low-cost SoC, especially IoT systems, which is cheap and compact and space-saving, and which can avoid the need for external components.

The object of the invention will be solved by a method for metering power consumption of a digital system-on-chip for predicting battery life of a battery used by the SoC, the method comprising the following steps:

- Partitioning of all digital circuitry of the SoC into

voltage and/or clock domains;

- Metering a power consumption of every domain by counting a charge pulse of each domain every time a gate of said domain has been switched;

- Accumulating said counted charge pulses of each domain by an accumulator;

- Calculating an overall sum R of all charges drained from the battery by said domains of the SoC;

- Post-processing said overall sum R for predicting battery life by an integrated micro-processor;

- Signaling an imminent battery loss.

The basic idea of the invention is to keep the numeric integration of the prior-art approach, but remove the series resistor and analog-to-digital converter (ADC) altogether and move the metering into the SoC. Instead, the current is determined by proxy of activity metering inside a SoC and post-processing the results with an integrated micro

processor .

Starting from the point that the power dissipation of a digital circuit in general is the sum of two main contributors: leakage power and dynamic power -

(eq. 1) . For battery life prediction, though, the charge

drained from the battery is the more important value: Q _tot =

(eq . 2 ) , and C _sw is the

parasitic capacitance that is recharged with every clock cycle. This capacitance, however, is not constant but depends on the actual circuit (e. g. clock gating domains) and even the actual values being processed (e. g. the output of an AND gate only changes when switching between a state where all inputs are ' 1' and at least one input is ' 0').

Finally, f is the current switching frequency. This value typically is not only changed over time (frequency scaling) , but also different between different parts of a SoC (clock domains) . During periods t with a constant voltage eq. 2 can be simplified to:

The dynamic part of eq. 3 can be further transformed to and using a clock domain specific effective capacitance C _eff;d instead of the actual value-dependent C _sw;d this can be further simplified to (eq. 5), whereas d indicates the respective domain.

Putting everything together, this leads to i. e. the total charge drained is the sum over all intervals ( volt_intv) with constant voltage U _v and the inner sum over all clock domains ( clk_domains) .

On an abstract level all digital circuitry of a SoC can be considered as communicating Mealy-automata . So, the SoC can be partitioned into voltage and/or clock domains, whereas a domain may be an individual clock domain within a more common voltage domain or a domain may be a voltage domain with a single clock domain or in principle a domain may be a voltage domain within a more common clock domain. However, this case is typically handled by breaking the common clock domain into two physical clock domains to ease physical implementation (it can still be considered the same logical clock domain, though) .

Each domain has its own gate count, may run different clock speeds and even different voltages. Nevertheless, it still boils down to a certain energy (or charge) per time-unit t that is used in a specific domain d.

In the sense of this invention a domain comprises a purely combinational logic of arbitrary complexity, a group of sequential elements (flip-flops) that store state between subsequent clock cycles, a clock gate, i.e. a circuit that switches a clock glitch-free on and off, based on a

condition, and a power switch. For the power metering procedure each clock domain comprises a N-to-1 divider that provides a counting pulse for the accumulation. The factor N _d is a compromise between accuracy, an averaging interval and the fact that the accumulation itself contributes to the power consumption (and the more frequent the accumulators are updated the more power the metering consumes itself) . In the extreme, N _d could be set to 1, but this would result in a higher power consumption of the metering and the

requirement for wider accumulators (since these would be updated more often) . The accumulator assigned to a domain accumulates the respective domain's charge value. The accumulators can be separated or integrated into their respective domains.

Partitioning of all digital circuitry of the SoC into voltage and/or clock domains is done at design time and mostly is necessary for backend/synthesis tasks, anyway.

In a preferred embodiment of the inventive method, the overall sum R of charges of all domains is calculated by whereas S _d corresponds to a domain scaling

factor and I _d corresponds to a domain-specific increment. In another preferred embodiment of the inventive method, I _d and S _d represents a combined proportionality factor of the domain and they depend on the gate-count of said domain, the switching activity of said domain and the voltage of said domain . In another further preferred embodiment of the inventive method, and since the gate-count of each domain is fixed at design time of the SoC and is thus constant, the switching activity is determined by netlist simulations at design time or by a calibration step of the SoC, which can be done e. g. during production test and measure the power consumption under known conditions (voltage levels and clock speeds) .

The gate-count is fixed at design time and is thus constant. The switching activity in general depends on the circuit's inputs (and clock frequency) and may be determined by netlist simulations at design time or only later by

calibration on real silicon.

In another further preferred embodiment of the inventive method, I _d and/or S _d are adjusted to a current voltage selection of the domain, whereas I _d and/or S _d are

reprogrammed dynamically by the integrated micro-processor or by an automatic voltage selector.

In the presence of voltage scaling, either the I _d or S _d must be adjusted according to the voltage. That and the potential calibration for the switching activity make it necessary to reprogram the I _d or S _d dynamically. If I _d is set to 1 and an integrated processor carries out the scaling and final summation, the reprogramming is easily done within the realm of that processor. Otherwise, the I _d or S _d could be

programmable registers in hardware or ROM-based lookup- tables to account for the voltage dependency. In case of automatic voltage scaling (i. e. not under software control) a hardware extension is necessary. In a preferred embodiment of the inventive method, nested clock domains are handled by identifying an appropriate level of clock domains, numerically integrating of their respective activity and averaging over all potential sub- domains . In today's SoCs it is typical that clock domains are nested. Nesting may be implemented explicitly by the designer or may be inferred automatically by synthesis tools. The inventive method can deal with nested clock domains in two possible ways: Firstly, it can instrument each sub-domain with the N- to-1 divider and accumulator structures, or secondly, it can treat sub clock domains simply as parts of their parent domain with just a reduced switching activity and adjust the I _d (or S _d appropriately) . The important step is to identify the appropriate level of clock domains for numeric

integration of their respective activity and averaging over all potential sub-domains. Too small a domain will unduly increase the metering overhead; too big a domain will impair the metering accuracy. The factors that correspond to the domain' s C _eff;d can be determined either by pre-silicon simulation (RTL or netlist) or characterization of the actual silicon. These factors, together with the desired accuracy/overhead compromise, will then be taken to define the level of clock domains at design time.

So, in an embodiment of the inventive method dealing with nested domains, the factors I and/or S _d are determined by netlist simulations at design time or by a calibration step of the SoC.

In a further preferred embodiment of the inventive method, a top accumulator is pre-set with a credit reflecting a battery capacity from which the accumulated charges of each domain are subtracted until said credit has expired and the battery has reached its end of life.

Instead of summing up the metered charges by the

accumulators shown above, it would also be possible to pre set the top accumulator with a "credit" that reflects the battery capacity (minus a reasonable safety margin) and then constantly subtract from that. When that credit has expired the battery has reached its end of life (or needs to be recharged) . This alternative approach is very easy to implement since software needs to know about the battery size to take appropriate action in the case of imminent battery loss.

The object of the invention will also be solved by an apparatus which is appropriate for performing the inventive method of metering power consumption of a SoC, especially an IoT UE . The inventive apparatus is part of the SoC and comprises all digital circuitry of the SoC, wherein said digital circuitry is divided into clock and/or voltage domains, whereas each domain is assigned to an accumulator for accumulating counted charge pulses of said domain, a collection means for calculating an overall sum R of all charges drained from the battery by said domains of the SoC and a micro-processor for post-processing said overall sum R for predicting battery life of the SoC.

In an embodiment of the inventive apparatus, each domain comprises combinational logic, a clock gate for switching a clock, storage means for storing a state between subsequent clock cycles and a N-to-1 divider for providing a counting pulse for the accumulation.

In each clock domain a N-to-1 divider is added that provides a counting pulse for the accumulation. The factor N _d is a compromise between accuracy, an averaging interval and the fact that the accumulation itself contributes to the power consumption (and the more frequent the accumulators are updated the more power the metering consumes itself) . In the extreme, N _d could be set to 1, but this would result in a higher power consumption of the metering and the requirement for wider accumulators (since these would be updated more often) . In another embodiment of the inventive apparatus, the accumulator comprises a domain-specific increment register. This register contains a domain-specific increment I _d indicating an individual characteristic of a respective domain.

In another further embodiment of the inventive apparatus, the calculation means comprises a domain scaling factor register. This means, that for the calculation a domain scaling factor register is used. This register contains a domain scaling factor S _d indicating the consumed charge per metering pulse.

In an embodiment of the inventive apparatus, the domain- specific increment register and/or the domain scaling factor register are programmable registers or ROM-based look-up- tables.

In a further embodiment of the inventive apparatus, each domain comprises a voltage selector for dynamic voltage scaling .

Summarizing, the inventive method and apparatus for metering power consumption of a digital SoC with the goal of

predicting battery life, has the following advantages over the current state-of-the-art:

- The apparatus perform the inventive method without any analog or mixed-signal circuitry, making it easy to integrate the procedure into a digital chip. It also does not require any discrete parts (like measurement

resistors) .

- The hardware requires only little gate count overhead to implement and adds virtually no cost. - The invention can be applied to a variety of SoCs, with virtually any gate count.

- The method described is independent of the actual silicon technology .

- The metering method works for both, rechargeable and non- rechargeable batteries.

- The method is compatible with dynamic voltage/frequency scaling approaches widely applied in the industry.

The presented approach needs to work around an intrinsic drawback it has against the known prior art method: With the charge/power metering sitting directly at the battery the prior-art method will automatically cover all contributors in a multi-chip platform (like e. g. a smart phone with radio, modem chip, application processor and display) . The novel and inventive method needs to collect metering

information from all individual digital chips instrumented as described in the invention and additionally model the contributions from non-digital parts (like e. g. a radio chip) . This, however, is less of a limitation as one would think: typically the software collecting all the metering information would control the platform and thus knows about the activity of these extra components. It is then only a matter of sufficiently exact power consumption

approximations for these extra components. The metering method presented here does not strive to achieve an

absolutely exact number of battery discharge, but merely a result with reasonable margin allowing predicting the end of battery life in the near future.

The invention will be explained in more detail using

exemplary embodiments. The appended drawings show

Fig. 1 State-of-the-art battery monitoring; Fig . 2 Illustration of partitioning into voltage/clock domains ; Fig. 3 Illustration of the inventive power metering with local accumulation;

Fig. 4 Extension for dynamic voltage scaling; Fig . 5 Embodiment of the inventive power metering with combined collection and accumulation; Fig. 6 Illustration of nested clock domains.

Figure 2 shows two domains 15 exemplifying the possible partitioning of the digital circuitry 9 of a system-on-chip (SoC) into domains. On an abstract level all digital

circuitry of a SoC can be considered as communicating Mealy- automata. A domain may be an individual clock domain within a more common voltage domain, or a domain may be a voltage domain with a single clock domain, or in principle a domain may be a voltage domain within a more common clock domain. However, this case is typically handled by breaking the common clock domain into two physical clock domains to ease physical implementation (it can still be considered the same logical clock domain, though) .

The symbols used in figure 2 have the following meaning:

The "clouds" represent purely combinational logic 10 of arbitrary complexity.

The AND-symbol marked with "CG" represents a clock gate 11, i. e. a circuit that switches a clock glitch-free on and off, based on a condition (typically a combination of a latch and an AND in physical implementation) .

The flip-flop 13 symbol stands proxy for a vector of

sequential (clocked) elements that store state between subsequent clock cycles.

The two switch symbols "SWu" and "SWd" represent power switching 12. It should be noted that in any real silicon only one of them is used. Whether it is switching against ground ("footer switches") or Vdd ("header switches") depends on the actual technology, but has no influence on this invention.

The vectors connecting the parts have the following meaning:

- i _x : Primary input vector of the domain.

- o _x : Primary output vector of the domain.

- n _x : State vector of the domain (identical to the number of sequential elements in the domain) .

- m _x : The feedback vector from state to next-state logic.

- k _x : The vector of primary inputs that contribute to the output vector through purely combinational logic.

- p _x : The cross-domain forwarding vector.

Each depicted domain 15 in figure 2 has its own gate count, may run at different clock speeds and even different

voltages. Nevertheless, it still boils down to a certain energy (or charge) per time-unit t that is used in a

specific domain d.

Figure 3 shows an embodiment of the present inventive power metering procedure translating eq. 6 into the requirement to numerically integrate (accumulate) a value for each domain that is proportional to its gate-count, the average

switching activity, and the square of the domain's voltage. Figures 3 and 5 show two potential embodiments. Figure 3 refers to the domains 15 from figure 2 but just depicts the important additions for power metering. In each clock domain 15 a N-to-1 divider 14 is added that provides a counting pulse for the accumulation 13.

The factor N _d is a compromise between accuracy, an averaging interval and the fact that the accumulation itself

contributes to the power consumption (and the more frequent the accumulators are updated the more power the metering consumes itself) . In the extreme, N _d could be set to 1, but this would result in a higher power consumption of the metering and the requirement for wider accumulators

(since these would be updated more often) .

The structures in figure 3 marked as "Accu A" and "Accu B" are the accumulators 16 for the respective domain's charge values. They are shown somewhat separated but could also be integrated into their respective domains 15. The structure marked as "Coll" 17 calculates the overall sum R of all domain's contributions. In essence, the circuitry of figure

3 calculates eq. 6 by proxy: (eq. 7) with R representing the result, S _d corresponding to the domain scaling factors Scale _A\B 20, and I _d corresponding to the domain-specific increments Incr _A\B 19. It should be noted that there is some freedom in implementation to calculate eq. 7. Either the I _d or the S _d can be set to 1. If I _d is set to 1 then this has the advantage that the local accumulators 16 can be implemented with fewer bits and only the

multipliers and final sum need to be wide. When S _d is set to 1 then the multipliers become simple wires, but the local accumulators need to be wider. Which implementation is chosen also depends on where the boundary between hardware and software is drawn. If the SoC has an integrated processor, which would often be the case, then this can take over the multiplication, collection and final accumulation of the result. In this case setting the I _d to 1 is

advantageous since this saves gates and the integrated processor typically has the necessary multiplier unit, anyway .

There is another aspect to the I _d and S _d , respectively: Since they represent the combined proportionality factor they depend on the domain's gate-count, the switching activity and the voltage. The gate-count is fixed at design time and is thus constant. The switching activity in general depends on the circuit's inputs (and clock frequency) and may be determined by netlist simulations at design time or only later by calibration on real silicon. In the presence of voltage scaling, either the I or S _d must be adjusted

according to the voltage. That and the potential calibration for the switching activity make it necessary to reprogram the I _d or S _d dynamically. If I _d is set to 1 and an integrated processor carries out the scaling and final summation, the reprogramming is easily done within the realm of that processor. Otherwise, the I _d or S _d could be programmable registers in hardware or ROM-based lookup-tables to account for the voltage dependency. In case of automatic voltage scaling (i. e. not under software control) a hardware extension as depicted in figure 4, is necessary. It shows essentially a section of figure 3 (one accumulator 16 and scaling multiplier 20) with the logic for the dynamic adjustment of the I _d or S _d based on the current voltage selection, depicted as multiplexers 21 selecting one value out of a set based on the current voltage selection. Figure 5 shows another, more hardware-centric embodiment to implement eq. 7. In this case the counting pulses from each domain 15 contribute immediately to the overall sum R using their individual increments I _d 19. Based on what was

explained previously, it should be clear that the I need to be dynamically adjustable in case of voltage scaling. This would either be done by changing the I _d through the

integrated processor or using the extensions shown in figure 4. It should be mentioned that simple voltage-switching does not require dynamic adjustment of the I _d , because during the off-time of a domain there are simply no count pulses. The embodiment according to figure 5 would typically still use a processor in the SoC to process the metered result and potentially notify the user of imminent battery loss.

It should be mentioned that for the scope of this invention an integrated processor of the SoC is also just a (set of) voltage/clock domain (s) that is metered along with all other domains. The complete metering needs a combination of hardware and software; hence the inventive apparatus is also called "hybrid power metering". The hardware would do the low-level metering on a per-domain basis during intervals with constant parameters (voltage, switching activity) .

Software is then responsible for integrating the domain' s results and would adjust the parameters accordingly. Since firmware would control the SoC power states (voltage

scaling, power-switching) anyway, it can easily apply the correct parameters for the respective metering intervals.

In another embodiment it is possible instead of summing up the metered charges by the accumulators 16 shown above, to pre-set the top accumulator with a "credit" that reflects the battery capacity (minus a reasonable safety margin) and then constantly subtract from that. When that credit has expired the battery has reached its end of life (or needs to be recharged) . This alternative approach is very easy to implement since software needs to know about the battery size to take appropriate action in the case of imminent battery loss.

It should be clear for the person skilled in the art that the embodiments according to figures 3 and figure 5 are but two possibilities out of a whole spectrum. It would be straightforward to apply figures 3 and figure 5 to different sections of a more complex SoC and then combine their results .

While figures 3 and figure 5 only show flat clock domains 15, it is more typical in today's SoC' s that clock domains are nested, as shown in figure 6. Nesting may be implemented explicitly by the designer or may be inferred automatically by synthesis tools. In figure 6 the top clock domain would drive all registers marked as "A". There are two direct sub- domains "B" and "C" that further condition the clocks (with individual conditions) . Domain "D" is yet one more sub-level of "C". The inventive method can deal with this situation in two possible ways: Firstly, it can instrument each sub- domain with the N-to-1 divider and accumulator structures shown in figures 3 and figure 5; or secondly, it can treat sub clock domains simply as parts of their parent domain with just a reduced switching activity and adjust the I _d (or S _d appropriately) .

The important step is to identify the appropriate level of clock domains for numeric integration of their respective activity and averaging over all potential sub-domains. Too small a domain will unduly increase the metering overhead; too big a domain will impair the metering accuracy. The factors that correspond to the domain' s C _eff;d can be

determined either by pre-silicon simulation (RTL or netlist) or characterization of the actual silicon.

Method and apparatus for metering power consumption of a digital SoC for predicting battery life

List of Reference Signs

1 Battery monitoring electronic

2 battery

3 charge electronic (CE)

4 series resistor

5 differential amplifier

6 analog-to-digital converter

7 numerical integration (NI) block

8 digital interface (DI)

9 functional circuitry

10 purely combinational logic

11 clock gate

12 power switch

13 flip-flop, storage means

14 N-to-1 divider with factor N _d

15 Respective domain

16 Accumulator

17 Collection means

18 Voltage selector

19 domain-specific increment register

20 domain scaling factor register

21 voltage multiplexer