Smartphone Thermal Analysis: Infrared Imaging Under Gaming and AI Workloads

Thermal design is now the primary limiter of smartphone performance. With 3nm SoCs packing 19–22 billion transistors and NPUs exceeding 45 TOPS, power density has outpaced the ability of a 75-gram glass sandwich to reject heat. Clock speed and core counts no longer tell the full story; the chassis is the heatsink.

In this analysis we used a calibrated FLIR A700 (7.5–14 µm, ±2°C, 640×480, 30Hz) with a blackbody reference, combined with on-device telemetry via kernel thermal zone logs and powerstats. Three current flagships (all vapor-chamber equipped, Snapdragon 8 Gen 4 / Dimensity 9400 class) were tested at 25°C ambient, 50% RH, on an open bench with no case, screen brightness locked to 400 nits. Each workload ran for 20 minutes or until Tskin stabilized.

Methodology: Beyond Point Sensors

Internal thermistors (typically NTCs near the SoC, battery, and 5G modem) report conservative, delayed junction-adjacent temperatures. They are excellent for safety but poor for spatial analysis. Infrared imaging gives us a continuous emissivity-corrected map of the entire rear surface, revealing lateral gradients, frame hotspots, and how effectively the cooling stack spreads heat away from the die.

We calibrated emissivity to 0.95 for matte glass and 0.88 for polished aluminum frames using Kapton tape references. All devices were matte-black skinned in the SoC region to avoid reflections, then validated against embedded Tskin thermocouples. Our error after correction was <1.1°C. Key metrics captured:

• Tskin_max: Peak rear surface temperature, typically 2–5mm offset from SoC center.
• ΔT_center-edge: Gradient from hotspot to mid-frame, indicating spreader efficiency.
• Time to throttle (TTT): Seconds until sustained power drops >10% from peak.
• Stabilized power (P_sus): Average SoC package power at thermal equilibrium (minutes 15–20).

Anatomy of the Modern Thermal Stack

A flagship thermal stack in 2026 is no longer just “copper and paste.” It is a system: SoC → vapor chamber (VC) → graphite spreaders → mid-frame → display and battery as thermal mass.

The key insight from the teardown work we covered in our Samsung Galaxy S26 Ultra teardown is scale: vapor chambers have grown from ~1,800 mm² to >4,200 mm² in two generations, now covering the SoC, RAM, and PMIC. The frame is no longer decorative — it is the heatsink fin.

Gaming Workload: 20 Minutes of Genshin Impact @ 60fps

We ran Genshin Impact (Sumeru city traversal, highest settings, 60fps cap) with a 5G SA connection and game mode enabled. This is a sustained GPU + CPU + modem load — typically 6.5–8.5W package power at launch.

All three devices showed an identical pattern: 45–80 seconds of full clocks, then a stepped decline. But infrared tells a more nuanced story than a FPS counter:

• Hotspot migration: In the first 90 seconds, peak Tskin sits directly over the SoC (≈ 38–39°C). By minute 8, the hotspot has bled into the upper third and the aluminum frame rails, which run 2–3°C hotter than the glass back due to higher conductivity and lower emissivity.
• Frame as radiator: Devices with a contiguous mid-frame-to-rail vapor chamber path showed a 9–11°C ΔT_center-edge, versus 14–16°C for graphite-only routing. That wider spread means lower peak, higher average — more comfortable in the hand, but warmer palms.
• Throttling signature: FPS did not drop smoothly. It fell in 3–4 discrete steps as the thermal daemon capped GPU from 1100 MHz → 850 → 680 MHz, correlating exactly with Tskin crossing 42.5°C, 43.5°C, and 44.2°C thresholds.

Device A maintained the highest sustained power because it could keep Tskin just under the 43.5°C mitigation trip. Device C, despite the same SoC, hit 45.2°C on a tight 22mm hotspot and was forced to throttle harder. Infrared confirmed Device C’s hotspot was 18% smaller in area — heat was trapped, not spread.

For gamers, the takeaway is tactile. The median palm contact area (index finger and middle finger on the rear, thumb on frame) overlapped the 42°C isotherm on Device C for 73% of the session versus 31% on Device A. That is the difference between “warm” and “uncomfortable” in a long session.

AI Workloads: The NPU Is a Different Beast

On-device AI presents a fundamentally different thermal problem. We tested three representative tasks: (1) Stable Diffusion 1.5 — 512×512 image generation (15 steps), (2) 7B LLM summarization of a 2,000-token document, and (3) continuous AI Super Resolution video upscaling (1080p → 4K @30fps).

Unlike gaming’s distributed GPU load, NPU workloads are dense and spiky. A single SD generation spikes to 10.2–11.8W for 7–9 seconds, then idles. The die’s NPU cluster is physically smaller than the GPU, so power density is 1.6–2.1× higher. IR frames captured a 4.5°C rise in the hotspot in just 6 seconds — faster than any gaming ramp.

LLM inference was most revealing. During the prefill phase (parallel token processing), power averaged 9.4W with a tight, circular hotspot centered on the SoC. During decode (autoregressive generation), power fell to 4.1W but remained on the NPU and memory controller, keeping Tskin elevated. Over five consecutive prompts, Tskin never recovered below 41°C; the device entered a “thermal ratchet” where each generation started hotter than the last. After the 4th prompt, Device B capped NPU frequency from 1.35 GHz to 920 MHz, increasing token latency from 28ms to 44ms.

This is where thermal analysis intersects with digital signal processing pipelines and modem workloads — the same power delivery network feeds the NPU, DSP, and 5G modem. Concurrent AI + network activity (e.g., live translation on a call) compounds the hotspot, which is why OEMs now gate NPU bursts when modem temp exceeds 43°C.

Telemetry Example: Catching the Throttle

Fusing IR data with kernel logs makes throttling deterministic, not anecdotal. Below is a simplified trace from Device A during LLM inference:

# fused log — t in seconds from start, Tskin from calibrated IR ROI over SoC
t=  0s  Ppkg=3.2W  Tskin=31.4°C  NPU=1350MHz  CPU_big=2.8GHz  prefill start
t=  6s  Ppkg=9.8W  Tskin=35.9°C  NPU=1350MHz  ΔTskin +4.5°C in 6s
t= 12s  Ppkg=9.4W  Tskin=39.1°C  NPU=1350MHz  thermal_zone9 (npu) = 98°C
t= 18s  Ppkg=4.2W  Tskin=40.8°C  NPU=1350MHz  decode phase, power drops but Tskin rises
t= 34s  Ppkg=4.0W  Tskin=41.9°C  NPU= 920MHz  // skin-therm throttle: Tskin >41.5°C, cdev8 state=1
t= 45s  Ppkg=3.6W  Tskin=42.3°C  NPU= 735MHz  // token latency 28ms → 41ms

The same mechanism that protects comfort degrades AI responsiveness. Users perceive it as “the phone got slow after a few AI generations,” not as heat.

Surface Temperature & Ergonomics: Where Hands Meet Heat

Regulatory touch limits (IEC 62368-1) allow 48°C for brief contact on glass, but comfort is far lower. We mapped hand contact zones using pressure film and overlaid them on IR isotherms. The critical zones are the upper rear quadrant (index finger rest) and the right frame rail (thumb grip in landscape).

After 15 minutes of gaming, the right rail on all devices exceeded 44°C on Device C and 42.5°C on Device A. In portrait AI use (one-handed, phone held low), the lower hotspot from the PMIC (which heats during fast charging + NPU) added a secondary 40–41°C zone near the USB-C port — often missed in reviews that only measure the SoC area.

Interestingly, UWB and secure element workloads like digital car key ranging show negligible thermal impact (<0.3W), but their antennas sit in the frame rails. Sustained high frame temperatures can slightly detune UWB ranging by 2–3cm due to thermal drift — a subtle coupling most thermal analyses ignore.

Practical guidance: a thin TPU case (0.8mm) adds ~1.2°C to Tskin (insulation) but reduces perceived temperature by 2–3°C due to lower thermal effusivity — it feels cooler even though the SoC is hotter and throttles ~7% sooner. Metal or no case feels hotter instantly but allows slightly longer sustained performance.

Mitigation: What Actually Works

Hardware and software co-design is the only path forward. Purely increasing VC area has diminishing returns beyond ~4000 mm² without improving the frame path. The best performing device in our set used a VC that is soldered (not just pressed) to the mid-frame and includes a graphite tongue that wraps around the battery, using it as a thermal capacitor for burst loads.

On the software side, OEM thermal daemons have evolved from simple frequency capping to workload-aware scheduling. The most effective strategies we observed:

• Skin-aware DVFS: Instead of waiting for junction temp, the daemon predicts Tskin 10–20 seconds ahead using a thermal RC model and pre-emptively shifts work from NPU to DSP or CPU efficiency cores.
• Spatial throttling: Migrating threads off the hottest core cluster, even at a small IPC cost, lowers hotspot intensity more than uniform downclocking.
• Burst pacing: For generative AI, inserting 400–600ms idle gaps between diffusion steps reduces peak Tskin by 1.8°C with only a 12% increase in total generation time — a worthwhile trade for comfort.

Users concerned about background thermal load from trackers and analytics should also consider that persistent background activity raises baseline Tskin by 0.5–1.0°C, shrinking thermal headroom. Our guide to mobile privacy and tracking prevention details how to audit wake-locks that contribute to this idle heat floor.

Conclusion: Measure What Your Hands Feel

Infrared imaging reframes performance: sustained watts are limited not by silicon, but by how evenly you can spread joules across a palm-sized surface without crossing 43°C. In our testing, the difference between the best and worst spreading design was 1.0W of sustained power and 10 fps — entirely from mechanical design, not SoC binning. As on-device AI becomes ambient and always-on, designing for tight, spiky NPU hotspots will matter more than designing for steady-state gaming. If you are benchmarking phones, demand 15-minute stabilized Tskin maps, not just 60-second benchmark scores. That is where real-world experience lives.