Temperature Sensor Chip
On-Die Temperature Sensors and PVT Monitors are the integrated measurement circuits distributed across the chip that continuously monitor die temperature, supply voltage, and process corner in real time — providing the feedback signals that thermal management systems, DVFS controllers, and reliability monitors need to keep the chip operating within safe bounds, where even a 10°C temperature error can lead to thermal throttling that wastes 15% performance or thermal runaway that damages the die.
Why On-Die Sensing
- External temperature: IR camera or thermocouple → slow, measures package not junction.
- On-die sensor: Directly at transistor level → measures actual junction temperature → fast.
- Modern chips: 10-50+ thermal sensors distributed across die → thermal map updated every 1-10 µs.
- Use: Dynamic thermal management (DTM), DVFS feedback, reliability monitoring.
Thermal Diode Sensor
- Most common: Forward-biased diode (substrate PNP BJT).
- Physics: VBE = (kT/q) × ln(IC/IS) → VBE is proportional to absolute temperature (PTAT).
- Measure VBE at two currents: ΔVBE = (kT/q) × ln(I₂/I₁) → temperature from voltage difference.
- Accuracy: ±1-3°C after calibration.
- Area: Very small (~100 µm²) → can place many across die.
PTAT (Proportional to Absolute Temperature)
- ΔVBE: Linear with temperature, process-independent → robust measurement.
- Combined PTAT + CTAT → bandgap reference (constant voltage) + temperature output.
Digital Temperature Sensor
| Architecture | Resolution | Conversion Time | Area | Power |
|---|---|---|---|---|
| BJT + Sigma-Delta ADC | 0.1°C | 10-100 µs | 0.01 mm² | 50-200 µW |
| Ring oscillator based | 0.5-1°C | 1-10 µs | 0.005 mm² | 10-50 µW |
| Time-to-digital (TDC) | 0.2°C | 5-50 µs | 0.008 mm² | 30-100 µW |
| All-digital (inverter delay) | 1-2°C | 0.1-1 µs | 0.002 mm² | 5-20 µW |
PVT Monitors
| Parameter | Sensor | What It Measures |
|---|---|---|
| Process (P) | Ring oscillator frequency | Fast/slow corner → actual transistor speed |
| Voltage (V) | Voltage divider + ADC | Local supply voltage at sensor |
| Temperature (T) | Thermal diode or RO | Local junction temperature |
- Ring oscillator: Frequency varies with PVT → combined indicator of actual circuit speed.
- Used for: Adaptive voltage scaling → measure actual speed → set minimum safe voltage.
- Critical path replica: Replica of worst critical path → directly measures timing margin.
Thermal Management Actions
| Temperature | Action | Response Time |
|---|---|---|
| < 85°C | Normal operation | — |
| 85-95°C | Reduce voltage (DVFS) | 10-100 µs |
| 95-105°C | Clock throttling | 1-10 µs |
| > 105°C | Emergency frequency reduction | Immediate |
| > 110°C | Thermal shutdown (THERMTRIP) | Hardware, < 1 µs |
Distribution Across Die
- CPU: 1-3 sensors per core + 1 per cache bank + 1 per memory controller.
- GPU: Sensor per SM cluster + per HBM PHY + per power rail.
- Total: 16-64 sensors on modern SoC → thermal map resolution ~1mm².
- Hotspot detection: Identifies which block is overheating → targeted throttling.
On-die temperature sensors and PVT monitors are the sensory nervous system of modern processors — without accurate, fast, distributed temperature and process monitoring, chips could not safely operate at the aggressive voltage and frequency points that deliver maximum performance, and the dynamic power management techniques that make modern mobile and server processors energy-efficient would be impossible.