on-die sensors

**On-die sensors** are **integrated measurement circuits** built directly on the semiconductor chip that monitor **temperature, voltage, process corner, and other physical parameters** in real time — providing the feedback data needed for adaptive power management, thermal protection, performance optimization, and reliability monitoring. **Why On-Die Sensors?** - External measurements (package temperature, board voltage) don't capture **within-die conditions** — hot spots, local IR drop, and process variation can only be seen from inside the chip. - Modern power management techniques (DVFS, AVS, ABB) require **real-time feedback** from the silicon itself. - **Thermal protection** requires knowing the actual junction temperature — not the ambient or package temperature. **Types of On-Die Sensors** - **Temperature Sensors**: Measure local junction temperature at specific die locations. - **BJT-Based**: Uses the temperature-dependent base-emitter voltage of a parasitic bipolar transistor. Most accurate (±1–2°C). - **Ring Oscillator-Based**: Frequency changes with temperature. Simpler but less accurate. - **Thermal Diode**: Forward voltage of a diode string changes linearly with temperature. - **Placement**: Multiple sensors distributed across the die — near CPU cores, GPU, memory controllers, I/O, and other hot spots. - **Voltage Sensors**: Measure local supply voltage to detect IR drop. - **ADC-Based**: Sample the local VDD and digitize it. Provides absolute voltage readings. - **Comparator-Based**: Compare local VDD against a reference — simpler, detects droop events. - **Purpose**: Identify IR drop hot spots, trigger DVFS adjustments, detect supply noise events. - **Process Monitors**: Determine the effective process corner of the local silicon. - **Ring Oscillators**: Frequency directly correlates with transistor speed — fast process = high frequency, slow process = low frequency. - **Leakage Monitors**: Measure standby current to determine effective $V_{th}$ — indicates fast/slow corner. - **Purpose**: Enable AVS and ABB — adjust voltage/bias based on actual silicon speed. - **Critical Path Monitors (CPMs)**: Replicas of actual timing-critical paths with delay measurement. - Track the actual timing margin of the design in real silicon. - More accurate than ring oscillators for predicting frequency capability. - **Aging Sensors**: Monitor degradation mechanisms. - **NBTI Monitors**: Track threshold voltage shift due to Negative Bias Temperature Instability. - **HCI Monitors**: Track Hot Carrier Injection degradation. - **Purpose**: Predict remaining lifetime, trigger compensating voltage adjustments. **Sensor Accuracy and Overhead** - **Area**: Each sensor typically occupies a small area (100–1000 µm²) — negligible for individual sensors but meaningful if hundreds are placed. - **Power**: Sensors consume small amounts of power — some can be duty-cycled (sampled periodically rather than continuously). - **Accuracy**: Temperature ±1–3°C, voltage ±5–10 mV — sufficient for management decisions. On-die sensors are the **eyes and ears** of modern chip power and thermal management — without them, the chip would operate blind, unable to adapt to its actual operating conditions.