Source-linked AI summary
Thermal-Headroom Routing for Efficient Robot Perception in Extreme Heat
Sparsh Roy, Kevin Zhou, Arhaan Shah, Shreyan Garg, Vihaan Joshi-Apte, Maxim Kudryashov, Nanda Cheruvanthala
TL;DR
Hot ambient conditions can make detector choices trigger thermal throttling, while latency-only routing may react after slowdown begins. THR uses a closed-form temperature forecast to route among detector configurations and avoid predicted throttling. In simulation, it reduced deadline misses and energy per on-time frame against latency-belief routing under tight deadlines, while tuned static rules could match it at fixed operating points.
Problem
On hot processors, slowdowns caused by earlier detector choices can persist for tens of seconds, so latency-only routing may react too late.
Source: §1 · pp. 1–2Method
THR uses a closed-form RC thermal forecast to exclude detector configurations predicted to cross the throttling threshold.
Source: §1 · p. 2Results
3.9–7.5 percentage-point fewer deadline misses and 12–13% lower energy per on-time frame than latency-only routing, with no throttling on all seven traces at 45 and 55 °C.
Source: §8 · p. 8Takeaways & Limitations
A tuned thermostat or duty cap can match THR at a fixed operating point; forecasting is most useful when ambient conditions or chip heating differ from static-policy assumptions.
Source: §7 · pp. 7–8, §8 · p. 8The study simulates throttling using a one-node RC plant; Apple M5 heating supports that approximation, but the M5 did not throttle, so hardware throttling remains unvalidated.
Source: §7 · p. 8
Abstract
from arXiv · showhide
Adaptive robot perception systems must balance detector accuracy, inference latency, and energy consumption, particularly when high ambient temperatures cause processors to thermally throttle. We propose Thermal-Headroom Routing (THR), a lightweight model-selection policy that forecasts processor temperature using a closed-form RC thermal model and excludes detector configurations predicted to cross a throttling threshold. We evaluate THR with four YOLO11 configurations across seven real-world perception traces, eight ambient temperatures, and 20 randomized seeds, comparing it with latency-aware routing, fixed-threshold thermostats, duty caps, and a matched-compute-budget control. At ambient temperatures of 45-55 degrees C and a tight inference deadline, THR eliminates throttling in simulation, reduces deadline misses by 3.9-7.5 percentage points, and reduces energy per on-time frame by approximately 12-13% relative to a latency-belief router. A separate experiment using deadline-paced, nonoverlapping frame execution yields miss-rate reductions of 6.2-9.5 percentage points. A tuned thermostat or duty cap can match THR at a fixed operating point; predictive routing is most useful when ambient conditions or calibrated thermal parameters differ from those used to tune static policies. Measurements on an Apple M5 support the single-node heating approximation but do not validate hardware throttling. The results characterize the conditions under which thermal forecasting improves the efficiency and timeliness of resource-constrained robot perception.
1. Introduction
Hot processors can make earlier detector choices cause prolonged slowdowns, so latency-reactive routing may respond too late. The paper introduces THR, a temperature-forecasting router, and evaluates when forecasting adds value over simpler rules.
1. Introduction
- At 45 °C, always running the large detector causes 12.5 s without an on-time detection, while a latency-reactive router enters repeated throttle cycles.The same always-large policy goes 91 s without an on-time detection at 55 °C.Source: §1 · p. 1
- Latency-reactive adaptive inference can respond too late because earlier choices cause slowdowns that persist for tens of seconds.Related approaches include early-exit networks, per-frame routers, and contention-aware detectors –.Source: §1 · pp. 1–2
- THR forecasts temperature with a closed-form rule and excludes configurations predicted to exceed the throttle limit, at about 7 µs per frame.The paper evaluates it against thermostats, offline-tuned duty caps, matched-budget control, deadline sweeps, and chip variations.Source: §1 · p. 2
- THR’s gain over latency-only routing comes from rationing the heavy model; forecasting helps when ambient conditions or chip heating differ from static-policy assumptions.With deadline slack, throttling can cost less than rationing does.Source: §1 · p. 2
2. Related Work
Prior work adapts inference to input, latency, or thermal behavior, but the paper distinguishes THR by letting detector choices affect the thermal state that routing must manage.
Adaptive and efficient inference
- Adaptive inference methods vary computation by input or latency, while ApproxDet selects detection branches under a latency SLA using measured contention.These approaches treat compute budgets as fixed or observed rather than as a state driven by policy choices.Source: §2 · p. 2
Thermal behaviour of edge inference
Thermal management
- Thermal-management methods predict hotspots or control frequency and placement, whereas THR leaves the governor alone and chooses among detector configurations using a temperature forecast,,,.The paper notes that THR could run on top of a DVFS policy such as zTT.Source: §2 · p. 2
Sensor- and compute-aware routing
- Belief-space perception routing [9] tracks blur, illumination, and occlusion faults per camera with switching HMMs, maintains a latency-based contention belief, and combines them via noisy-OR; this work reuses its fault beliefs, frontier, and traces from TartanDrive,, RADIATE [22], and WoodScape.Source: §2 · p. 2
3. System Model
The system model couples detector choices to processor temperature and throttling, then evaluates routing on trace frames under a specified deadline and timing protocol.
Frontier and deadline
- The system uses four detector configurations, with the base deadline set to the largest model’s median normal-compute latency; accuracy is agreement with that model’s detections.This pseudo-ground truth ranks configurations differently from real annotations, while miss, throttling, and energy numbers do not depend on it.Source: §3 · p. 2
Thermal plant
- The processor is modeled as one thermal node with time constant τ and junction-to-ambient resistance Rth, and configuration-dependent power drives its temperature.The chosen plant represents a sealed, sun-loaded enclosure; continuous C1 operation settles at ambient plus 50 °C, so throttling starts above 40 °C ambient.Source: §3 · p. 2
- Throttling depends on prior routing choices, so policies run in closed loop with shared pre-drawn latency samples and sensor noise for each seed.While throttled, every configuration uses its profiled contended latency distribution, whose p95 is about 2.5× normal.Source: §3 · p. 2
Frame timing
- Main experiments play traces at their 10 Hz recording rate, drop late frames without queuing, and advance the thermal model by 0.1 s per decision.A separate heat experiment uses physically consistent timing, with each decision lasting exactly D.Source: §3 · pp. 2–3
4. Policies
All policies select among configurations using predicted compute state, deadline feasibility, and expected accuracy; THR adds a forecast-based thermal constraint.
4. Policies
- Policies predict compute state, retain configurations whose median latency meets the deadline, and serve the highest-expected-accuracy option given camera-fault belief.A 10-frame dwell prevents rapid switching.Source: §4 · p. 3
- THR forecasts the temperature after holding a configuration for 50 frames and drops it when the forecast reaches the throttle limit minus a 2 °C margin.The forecast uses the on-die temperature reading and a closed-form thermal model; evaluating it costs 7.3 µs per frame on one Apple M5 core.Source: §4 · p. 3
- Baselines include latency-only routing, a thermostat that bars C1 at 88 °C, duty caps tuned to THR’s C1 share, and a matched-budget control using that same share.The matched-budget control tests whether THR helps by timing C1 frames or only by using fewer of them.Source: §4 · p. 3
5. Experimental Setup
The evaluation uses seven real-world perception traces across a set of thermal and routing experiments. Its 20-seed protocol fixes headline settings before evaluation and pairs policy comparisons on shared randomized inputs.
Traces
- Seven traces, including TartanDrive, RADIATE, and WoodScape footage, were repeated to fill 1,200-decision missions.The set includes a no-camera-fault negative control, rough terrain, weather conditions, night driving, and real soiling masks.Source: §5 · p. 3
Experiments
- Experiments varied ambient temperature, contention, deadline, chip parameters, thermal-model inputs, and policy thresholds, with a separate run using physically consistent frame timing.E2 sweeps 20–55 °C; E11 repeats it with each decision lasting exactly the base deadline.Source: §5 · p. 3
Protocol
- Headline settings were fixed before evaluation; 20 seeds varied latency, camera noise, and temperature-sensor noise on the same real frames.Comparisons use paired per-seed t-tests, with Holm correction across seven traces; caps tuned at 45 and 55 °C were fit on evaluation seeds, favoring them.Source: §5 · p. 3
6. Results
At hot ambient temperatures, THR avoids simulated throttling and improves timing and energy relative to Belief, but incurs an accuracy trade-off. Static rules can match it at a fixed operating point, while their performance can shift across temperatures.
6.1. Cool air
- At 35 °C and below, no policy reached the throttle threshold, and the principal routers made identical choices with miss rates differing by 0.0 points across all seven traces.At 40 °C, THR held C1 to 93–94% of frames because its 5 s forecast crossed 88 °C.Source: §6.1 · p. 3
6.2. Hot air, against latency-only routing
- Belief and Reactive had roughly 1 s blackouts but were throttled for 11–13% of the mission at 45 °C and 32–38% at 55 °C.Their latency response follows the governor stepping in; Static C1 was throttled 20% at 45 °C and 76% at 55 °C.Source: §6.2 · p. 3
- THR was never throttled and reduced deadline misses versus Belief by 3.9–5.4 points at 45 °C and 6.7–7.5 at 55 °C; energy per on-time frame fell 12% and 13%, respectively.Mean agreement at 45 °C was 0.950 for THR versus 0.975 for Belief; on-time utility rose significantly on 5 of 7 traces at 45 °C and 6 of 7 at 55 °C.Source: §6.2 · p. 3, Table 2 · p. 4
6.3. Simple temperature rules and the trade-off curve
- The 88 °C thermostat also avoided throttling, but THR missed 1.3–1.9 fewer points at 45 °C and 2.9–4.4 fewer at 55 °C, while retaining less accuracy.The thermostat’s threshold sweep traces an accuracy–timeliness curve on which THR also falls.Source: §6.3 · p. 4
- At one ambient on the assumed chip, matched budget matched THR’s miss rate on all seven traces at 45 °C and six of seven at 55 °C, showing that C1 share largely determines the outcome.The matched-budget control uses THR’s C1 share without temperature input.Source: §6.3 · p. 4
- A cap tuned at 45 °C throttled 16% of the mission at 50 °C and 31% at 55 °C, while THR missed 3.1–6.6 fewer points; below 45 °C, the cap gave up 3.6–5.4 points of accuracy.A cap tuned at 55 °C cost 7.1 points of accuracy at 45 °C and 12.1 at 35 °C.Source: §6.3 · p. 4
6.4. A different chip, with a calibrated forecast
With the matching thermal model, THR stayed unthrottled across the tested chip variants; under deadline-paced execution it also reduced misses and energy use relative to Belief. The comparison tests a calibrated forecast against static policies tuned for a different plant or operating point.
6.4. A different chip, with a calibrated forecast
- THR stayed at 0.0% throttling across all five chip variants, temperatures, and traces when given the matching model.The 88 °C thermostat throttled up to 29.0% of the mission at 55 °C when Rth was 20% higher, and up to 2.7% when τ was halved.Source: §6.4 · p. 4, Table 5 · p. 6
- The 88 °C thermostat can reach 90 °C after crossing its threshold, while THR forecasts five seconds ahead and detects the overshoot.On the slow-heating chip at 45 °C, THR’s 3.7–4.4-point miss gain came from rationing C1 despite no Belief throttling.Source: §6.4 · p. 4
- The comparison does not test THR on hotter chips at 55 °C with an uncalibrated model, and no retuned thermostat was evaluated.A seven-minute heating curve supplied the M5’s Rth and τ; a thermostat threshold depends on both chip and ambient.Source: §6.4 · p. 5
6.5. Physically consistent frame timing
- The E11 evaluation uses one decision per deadline period, skips intervening trace frames, abandons unfinished inference, and keeps THR’s horizon at five seconds.Missions last 1,200 decisions, or 4.2–6.8 minutes; the 45 °C cap was retuned for this timing.Source: §6.5 · p. 5
- 6.2–9.5 percentage points fewer deadlines were missed by THR than Belief across 45–55 °C, with 15–18% less energy per on-time frame.THR was never throttled; the matched budget tied THR on misses at every ambient.Source: §6.5 · p. 5, Table 6 · p. 6
6.6. When the deadline has slack
With deadline slack, THR’s miss-rate advantage over Belief was small, while Belief sometimes had higher on-time utility. Across fault-coupled contention and thermal-model settings, the results show distinct contributions from fault handling and thermal forecasting, with burst-level benefits more limited.
6.6. When the deadline has slack
- At 1.25× and 1.5× the median deadline, THR missed 0.0% of frames while Belief missed 0.5% at 45 °C and 1.0% at 55 °C.Belief had higher on-time utility on six of seven traces (0.970 versus 0.950 at 45 °C), while THR used 5% less energy per on-time frame (2.17 versus 2.28 J at 45 °C).Source: §6.6 · p. 6, Table 7 · p. 7
6.7. Heat combined with camera faults
- At 45 °C and 55 °C, Joint + THR beat Belief on every trace by 5.8–13.1 and 7.3–13.3 points, respectively, and never underperformed Joint alone.On the fault-free drive, Joint gained 0.0 points while THR gained 6.0 points at 45 °C and 7.9 at 55 °C; at 25 °C THR added nothing.Source: §6.7 · p. 6, Table 8 · p. 7
6.8. THR’s settings and a wrong model
- Across the tested horizons and margins, THR throttling remained at 0.0% except for 0.6% under H = 20 and m = 0 °C at 55 °C.At 45 °C, miss reduction over Belief ranged from 0.5 to 11.0 points; underestimating Rth by 30% reduced the gain to 0.8–1.8 points.Source: §6.8 · p. 6, Table 9 · p. 7
6.9. Bursts of misses under fault-coupled contention
- Fault coupling reduced the share of frames in miss bursts of at least three on six of seven traces by 0.6–5.6 points, with no change on the clean drive.It significantly shortened the longest burst on only one trace.Source: §6.9 · p. 6
6.10. Heating measured on a real chip
Apple M5 measurements support a single-node RC approximation for heating under the tested YOLO11 workloads. The laptop did not throttle, so the measurements do not validate hardware throttling.
6.10. Heating measured on a real chip
- The M5 test ran C1, C2, and C4 nonstop for seven minutes on 1024 × 544 frames, fitting the step response to GPU temperature.GPU and CPU temperature, power, and fan speed were recorded four times per second.Source: §6.10 · p. 7, Table 11 · p. 7
- A single RC node fit the measured heating curves with under 1 °C error, but fitted parameters varied by workload.For C1, the fitted time constant was 20 s and resistance 2.8 °C/W; for C4, they were 45 s and 3.3 °C/W.Source: §6.10 · p. 7
- The Apple M5 did not throttle: C1 latency was 187 ms in the first 30 seconds and 186 ms in the last 30 seconds.The measurements support the heating model and parameter scale, but the fan-cooled laptop avoided reproducing throttling.Source: §6.10 · p. 7
7. Discussion and Limitations
Temperature forecasting keeps THR out of throttling when latency-only routers do not, but a tuned thermostat or duty cap can match it at a fixed operating point. The value of forecasting depends on deadline slack and on how well static policies reflect the chip’s thermal behavior.
7. Discussion and Limitations
- At a fixed operating point, a tuned thermostat or duty cap matches THR, while forecasting helps when ambient conditions or calibrated thermal parameters differ from those used to tune static policies.In those differing conditions, the thermostat and fixed caps throttled while THR did not.Source: §7 · pp. 7–8
- With 25% deadline slack, cheaper models absorb throttling and latency-only routing delivers more on-time accuracy than THR; deployments should check their deadline regime.At the large model’s median-latency deadline, avoiding throttling reduces misses and energy.Source: §7 · p. 8
- The evaluation is a simulation using real camera traces and profiled latencies; Apple M5 heating fits the one-node RC model, but the M5 did not throttle.The throttle behavior is therefore not validated on hardware; the authors identify hardware measurements as necessary.Source: §7 · p. 8, Figure 5 · p. 8
- Planned work will measure thermal parameters and throttle behavior on a Jetson Orin in an environmental chamber from 25 to 55 °C.The authors also plan to score routers against real annotations and test whether an online-learned duty budget can replace the RC model.Source: §7 · p. 8
8. Conclusion
THR forecasts temperature to avoid throttling and improve timeliness and energy efficiency under tight deadlines. Its advantage is greatest when thermal behavior differs from the assumptions behind static policies, while deadline slack can make its accuracy cost unwarranted.
8. Conclusion
- Across seven traces at 45 and 55 °C with a tight deadline, THR kept the chip out of throttling and cut deadline misses by 3.9 to 7.5 percentage points and energy per on-time frame by 12 to 13% versus latency-only routing.The temperature forecast costs microseconds per frame.Source: §8 · p. 8
- On a chip 20% hotter than assumed, the tuned thermostat throttled for up to 29% of the mission while THR stayed at zero.A tuned thermostat reaches the same operating points on the chip for which it was tuned.Source: §8 · p. 8
- With deadline slack, throttling is cheap and THR’s rationing is not worth its accuracy cost; the practical advice is to use a forecast when chip thermal behavior is uncertain.Source: §8 · p. 8