57  Edge Review: Power Optimization

In 60 Seconds

Deep sleep mode (0.01 mA) extends IoT battery life from 94 days to over 1.4 years compared to regular sleep (1 mA), a 5.6x improvement that makes the 2-second wake-up penalty negligible for devices sampling every 10+ minutes. For a 1000-device deployment over 5 years, optimized deep sleep saves over $400,000 in battery replacement costs.

57.1 Learning Objectives

By the end of this chapter, you will be able to:

  • Calculate Battery Life: Compute expected battery duration for various power profiles and duty cycles
  • Evaluate Sleep Mode Trade-offs: Compare deep sleep vs regular sleep including wake-up penalties
  • Design Duty Cycling Strategies: Determine optimal sampling intervals for different applications
  • Quantify Cost Savings: Calculate battery replacement savings from power optimization

Key Concepts

  • Power consumption profile: A characterisation of a device’s energy draw across all operating modes (active, idle, sleep, transmit) including the time spent in each mode during normal operation.
  • Energy budget: The maximum energy available to an edge device per unit time (from battery or harvesting), constraining the sum of all activities (sensing, processing, communication) to a sustainable average power.
  • Transmission power scaling: Adjusting radio transmit power to the minimum needed to achieve reliable communication with the gateway, saving energy on links that are shorter than the radio’s maximum range.
  • Adaptive sampling for power reduction: Reducing sensor sampling rate during quiet periods and increasing it when signals are changing rapidly, balancing data fidelity against energy consumption dynamically.
  • Power profiling tool: Hardware or software instruments (current probes, power monitors like Nordic PPK2, Nordic Energy Profiler) used to measure actual device current consumption across operating modes.

57.2 Prerequisites

Before studying this chapter, complete:

IoT devices are like hibernating bears:

  • Active mode (25 mA): Bear is hunting - uses lots of energy
  • Transmit mode (120 mA): Bear is running - uses MOST energy
  • Sleep mode (1 mA): Bear is napping - uses some energy
  • Deep sleep mode (0.01 mA): Bear is hibernating - uses almost no energy

The trick is keeping the bear in deep hibernation as much as possible, only waking it briefly to check things and report back.

A 2500 mAh battery is like 2500 hours worth of 1 mA activity. If your device averages 0.2 mA, it lasts 12,500 hours (over 1.4 years).

57.3 Power Profile Fundamentals

57.3.1 Typical IoT Device Power States

State Current Draw Duration Activity
Deep Sleep 0.01 mA 99% of time RTC running, RAM retained
Sleep 1 mA N/A Peripherals standby
Active 25 mA Milliseconds Sensing, processing
Transmit 120 mA Milliseconds Radio transmission

57.3.2 The Power Budget Equation

The average current over a duty cycle determines battery life:

\[I_{\text{avg}} = \frac{I_{\text{active}} \times t_{\text{active}} + I_{\text{tx}} \times t_{\text{tx}} + I_{\text{sleep}} \times t_{\text{sleep}}}{t_{\text{cycle}}}\]

\[\text{Battery Life (hours)} = \frac{\text{Battery Capacity (mAh)}}{I_{\text{avg}} \text{ (mA)}}\]

57.4 Deep Sleep Trade-off Analysis

The key question: Is the wake-up penalty worth the sleep savings?

## Deep Sleep Decision Matrix

57.4.1 When to Use Deep Sleep

Condition Recommendation
Long sleep intervals (>10 minutes) Use deep sleep
Low duty cycle (<1% active) Use deep sleep
Remote deployments Use deep sleep
Battery replacement costly Use deep sleep

57.4.2 When NOT to Use Deep Sleep

Condition Recommendation
Frequent wake-ups (<1 minute) Use regular sleep
High duty cycle (>10% active) Active current dominates anyway
Low-latency requirements 2-second wake-up too slow
Always-on peripherals needed Sleep mode retains peripherals

57.5 Battery Life Comparison

57.5.1 Scenario: Environmental Monitor

Parameter Value
Battery capacity 2500 mAh
Sampling interval 10 minutes
Active duration 0.1 seconds
Transmit duration 0.5 seconds

57.5.2 Cost Impact

The battery life extension from deep sleep mode demonstrates the power of low-duty-cycle operation:

\[\text{Battery Life (hours)} = \frac{\text{Battery Capacity (mAh)}}{\text{Average Current (mA)}}\]

For a 2500 mAh battery with 10-minute sampling intervals:

Sleep mode calculation:

  • Active: 0.1 sec at 25 mA = 2.5 mA-sec
  • Transmit: 0.5 sec at 120 mA = 60 mA-sec
  • Sleep: 599.4 sec at 1 mA = 599.4 mA-sec
  • Total: \((2.5 + 60 + 599.4) / 600 = 1.103\) mA average
  • Battery life: \(\frac{2500}{1.103} = 2{,}266 \text{ hours} \approx 94 \text{ days}\)

Deep sleep mode calculation:

  • Wake-up: 2 sec at 25 mA = 50 mA-sec
  • Active: 0.1 sec at 25 mA = 2.5 mA-sec
  • Transmit: 0.5 sec at 120 mA = 60 mA-sec
  • Deep sleep: 597.4 sec at 0.01 mA = 5.97 mA-sec
  • Total: \((50 + 2.5 + 60 + 5.97) / 600 = 0.197\) mA average
  • Battery life: \(\frac{2500}{0.197} = 12{,}690 \text{ hours} \approx 1.45 \text{ years}\)

\[\text{Life Extension} = \frac{I_{\text{sleep}}}{I_{\text{deep}}} = \frac{1.103}{0.197} \approx 5.6\times \text{ improvement}\]

Fleet-scale savings (1,000 devices over 5 years):

  • Sleep mode: 3.9 changes/yr per device \(\times\) 1,000 \(\times\) 5 = 19,500 changes at $25 = $487,500
  • Deep sleep: 0.69 changes/yr per device \(\times\) 1,000 \(\times\) 5 = 3,450 changes at $25 = $86,250
  • Savings: $401,250

The 2-second wake-up penalty adds only 50 mA-sec to a 600-second cycle – less than 0.1 mA to the average.

57.6 Priority-Based Processing for Mixed Deployments

Industrial facilities often have mixed requirements: some sensors are safety-critical, others are monitoring-only.

### Dual-Path Processing Architecture

Architecture diagram with two processing paths: a teal Critical Path box for safety sensors connected to an orange Reserved CPU box with guaranteed processing, and a blue Normal Path box for monitoring sensors connected to a purple Best Effort box with adaptive sampling, all flowing from an input sensor array on the left to output actions on the right.
Figure 57.1: Dual-path edge processing architecture showing critical safety sensors with reserved CPU allocation on a priority path and monitoring sensors on a best-effort path with adaptive sampling.

57.6.1 Adaptive Sampling Under Load

The system adjusts sampling rates based on CPU utilization:

CPU Load Data Retention Rationale
Low (<50%) Keep 100% of data Capacity available for all sensors
Medium (50-70%) Keep 90% of data Minor shedding of redundant readings
High (70-85%) Keep 80% of data Per non-critical sensor tolerance
Overload (>85%) Keep 70% of data Maximum acceptable data loss

57.7 Event-Driven Architecture Benefits

Comparing polling vs event-driven approaches for a 2500 mAh battery:

Approach Description Avg Current Battery Life
High-frequency polling Sample at 10 Hz, always active ~12 mA ~9 days
Periodic polling Sample every 10 min, sleep between ~1.1 mA ~94 days
Event-driven + deep sleep Wake on threshold events only ~0.2 mA ~1.4 years
Optimized event-driven Minimal wake, low-power comparators ~0.05 mA ~5.7 years

Event-driven architectures with deep sleep can achieve battery lifetimes exceeding a year, compared to just 9 days for high-frequency polling.

Scenario: Biologists deploy GPS collars on 100 elk to study migration patterns. Collars must last 2 years without battery replacement.

Initial Design (Fails Requirement):

  • GPS sampling: Every 5 minutes (288 readings/day)
  • Cellular transmission: Immediate upload after each reading
  • Battery: 6,000 mAh
  • Power profile:
    • GPS fix: 50 mA for 30 seconds (cold start)
    • Cellular TX: 200 mA for 5 seconds
    • Sleep: 2 mA between readings
  • Cycle time: 300 seconds (5 minutes)
  • Average current: \((50 \times 30 + 200 \times 5 + 2 \times 265) / 300 = (1500 + 1000 + 530) / 300 = 10.1\) mA
  • Battery life: 6,000 / 10.1 = 594 hours = 24.7 days (fails 2-year requirement)

Optimized Design:

Change 1: Event-Driven GPS

  • Only activate GPS if accelerometer detects movement
  • Elk stationary 70% of time (grazing, resting)
  • GPS activations reduced to 30% of readings

Change 2: Hot-Start GPS

  • Cache satellite ephemeris data for hot-start fixes
  • GPS fix time reduced from 30 seconds to 3 seconds
  • Power per fix: 50 mA for 3 seconds = 150 mA-sec (vs 1,500 mA-sec)

Change 3: LoRaWAN Instead of Cellular + Daily Batch Upload

  • Store GPS fixes locally, upload once per day
  • LoRaWAN: 40 mA for 1 second per packet
  • Daily upload: 12 packets at 1 second each = 12 seconds/day

Change 4: Deep Sleep

  • Replace 2 mA sleep with 0.01 mA deep sleep

New Power Calculation:

Per day (86,400 seconds):

  • GPS fixes: 288 readings \(\times\) 0.3 (moving) \(\times\) 50 mA \(\times\) 3s = 12,960 mA-sec
  • LoRaWAN TX: 12s \(\times\) 40 mA = 480 mA-sec
  • Deep sleep: ~86,376s \(\times\) 0.01 mA = 864 mA-sec
  • Total: 14,304 mA-sec per day = 3.97 mAh/day

Battery life: 6,000 mAh / 3.97 mAh/day = 1,511 days = 4.1 years (exceeds 2-year requirement)

Cost Savings:

Approach Battery Life Collar Replacements/2yr Cost @ $150/replacement
Original 24.7 days 100 collars x 30 replacements = 3,000 $450,000
Optimized 4.1 years 0 (no replacement needed) $0
Savings - - $450,000 over 2 years

Additional benefits:

  • No capture/recollaring stress to animals
  • Continuous multi-year migration data
  • Remote collar activation/deactivation via LoRaWAN downlink
Optimization Battery Life Gain Latency Cost Data Quality Cost When to Use
Deep sleep (0.01 mA) 5-10x +2s wake-up None Sampling interval >5 min
Event-driven sampling 3-10x None (faster) May miss slow changes Threshold-based monitoring
Batch transmission 5-20x Hours (deferred) None Non-real-time applications
Lower-power radio 3-5x None Shorter range LoRa vs cellular
Reduce sampling rate Linear with reduction None Lower resolution Slow-changing phenomena
Local processing 2-10x (avoid TX) None Must have edge compute High data volume

Selection guidance:

  • Need <2x improvement: Reduce sampling rate
  • Need 2-10x improvement: Add deep sleep and reduce sampling
  • Need 10-50x improvement: Combine deep sleep, event-driven sensing, and batch transmission
  • Need >50x improvement: All optimizations plus energy harvesting (solar, vibration)
Common Mistake: Forgetting Always-On Sensor Power in Event-Driven Designs

The Mistake: Students assume event-driven sensing eliminates all active power, forgetting that the accelerometer or comparator itself consumes power continuously.

Example Calculation Error:

Student calculates:

  • “Deep sleep 99.9% of time at 0.01 mA = 0.01 mA average”
  • Forgets: Accelerometer running 24/7 at 0.2 mA

Actual average current: 0.01 + 0.2 = 0.21 mA (21x higher than calculated)

Battery life impact:

  • Calculated: 6,000 / 0.01 = 600,000 hours = 68.5 years
  • Actual: 6,000 / 0.21 = 28,571 hours = 3.26 years

The Lesson: Event-driven systems still need always-on sensors or comparators. Include their power consumption in calculations. Choose ultra-low-power accelerometers (<0.05 mA) or use external hardware wake interrupts where possible.

57.8 Chapter Summary

  • Deep sleep mode (0.01 mA) provides a 5.6x power reduction over regular sleep (1 mA) for a typical 10-minute sampling cycle, extending battery life from 94 days to 1.45 years with a 2500 mAh battery.

  • Battery life calculations must account for all phases: wake-up penalty, active sensing, transmission, and sleep current. The wake-up penalty (50 mA-sec per cycle) is negligible compared to the sleep savings (from 599 mA-sec to 5.97 mA-sec per cycle).

  • Dual-path processing separates critical safety sensors (deterministic latency, zero data loss) from monitoring sensors (best-effort, adaptive sampling) to optimize mixed deployments.

  • Event-driven architectures with deep sleep achieve over a year of battery life compared to 9 days for high-frequency polling, but designers must account for always-on sensor power (accelerometers, comparators).

  • The decision matrix guides sleep mode selection: use deep sleep for long intervals, low duty cycles, and remote deployments; use regular sleep for frequent wake-ups or low-latency requirements.

Key Takeaway

Deep sleep mode is the single most impactful power optimization for low-duty-cycle IoT devices. At 0.01 mA (compared to 1 mA for regular sleep and 120 mA for transmitting), deep sleep extends battery life from 94 days to over a year for devices sampling every 10 minutes. The 2-second wake-up penalty adds only 0.083 mA to the average current across a 600-second cycle. For mixed deployments, dual-path processing separates safety-critical sensors (zero data loss, under 100ms latency) from monitoring sensors (best-effort, adaptive sampling). Always include always-on sensor power in event-driven designs.

“Bella’s Big Sleep!”

Bella the Battery was worried. “At this rate, I’ll be completely empty in just 94 days! That’s only about three months!”

“What’s using all your energy?” asked Sammy the Sensor.

“The sleep mode! Even when nobody’s doing anything, the device is still using 1 milliamp of power. It’s like leaving a night light on 24 hours a day.”

Max the Microcontroller had an idea. “What if we use DEEP sleep instead? It’s like hibernation! Instead of 1 milliamp, we’d only use 0.01 milliamp. That’s 100 times less!”

“But there’s a catch,” Lila the LED warned. “Waking up from deep sleep takes 2 whole seconds. Regular sleep only takes 0.1 seconds.”

“That’s okay!” Max calculated quickly. “We only check the temperature every 10 minutes. So we’re asleep for 597 seconds and awake for about 3 seconds. The wake-up penalty is tiny compared to all that sleeping time!”

Bella did the math. “Wait… instead of 94 days, I’d last over a YEAR AND A HALF?!”

“That’s right – more than five times longer!” Max said. “And if we only wake up when something important happens, like a temperature spike, we can save even MORE energy!”

Bella beamed. “Deep sleep is my new best friend! It’s like the difference between a cat nap and full-on bear hibernation!”

The Sensor Squad learned: For IoT devices that don’t need to check things very often, deep sleep is like a superpower. A tiny wake-up delay is a tiny price to pay for over a year of extra battery life!

57.9 Concept Relationships

Power optimization builds on:

  • Edge Architecture - Level 1 (device) power states and Level 3 (gateway) bundling strategies
  • Data Reduction - Bundling 60 transmissions into 1 reduces power consumption by reducing radio-on time

Power optimization enables:

  • Edge Deployments - Deep sleep extends battery life from days to years, enabling remote deployments
  • Storage Economics - Reduced battery replacement costs (saving over $400,000 over 5 years for 1,000 devices)

Parallel concepts:

  • Deep sleep trade-off analysis and edge vs cloud decision framework: Both balance competing constraints (wake-up penalty vs power savings; latency vs capability)
  • Dual-path processing and tiered storage: Both separate critical (real-time) from non-critical (best-effort) workloads

57.10 See Also

Related review chapters:

Foundational chapters:

Next topic:

57.11 What’s Next

Direction Chapter Link
Next Edge Review: Storage and Economics edge-review-storage-economics.html
Previous Edge Review: Gateway and Security edge-review-gateway-security.html
Related Edge Review: Architecture and Reference Model edge-review-architecture.html
Related Edge Review: Data Reduction Calculations edge-review-data-reduction.html

Common Pitfalls

For devices with long sleep periods and short active bursts, active current dominates energy consumption despite the short duration. Use power profiling to identify whether sleep optimisation or active-phase optimisation delivers more benefit.

Datasheet sleep currents assume all peripherals are properly disabled and no floating inputs are drawing current. Achieving datasheet sleep current in real hardware requires careful board design and software configuration.

A radio transmission that fails and requires 3 retries consumes 4× the expected transmission energy. Design communication protocols to minimise retransmissions through appropriate retry limits, backoff algorithms, and link quality monitoring.