kT/C Noise

kT/C noise arises from the thermal agitation of charge carriers in the non-zero resistance of a sampling switch. During the sampling phase, this resistance generates a broadband Johnson–Nyquist noise voltage. When the switch opens, a random noise sample is frozen onto the capacitor, creating an unpredictable charge error that is independent of the signal. This error is a physical manifestation of the fluctuation-dissipation theorem and cannot be eliminated by circuit design improvements alone.

The total mean-squared noise voltage on the capacitor is determined solely by the absolute temperature (T) and the capacitance value (C). Key relationships:

• RMS noise voltage: v_n = √(kT/C)
• Noise charge: Q_n = √(kTC)
• Noise energy: E_n = ½kT This reveals that the noise is independent of the switch resistance (R), as a larger R increases noise spectral density but reduces bandwidth proportionally. For a 1 pF capacitor at 300 K, the RMS noise voltage is approximately 64.4 µV.

The noise amplitude is inversely proportional to the square root of capacitance, creating critical design trade-offs:

• Doubling C reduces RMS noise by only √2 (≈30%), while doubling power consumption for charging
• Halving C increases noise by √2 but reduces area and power
• Practical range: 0.1 pF (≈204 µV RMS) to 100 pF (≈6.4 µV RMS) at room temperature This forces designers to balance noise performance, power dissipation, and sampling speed, making kT/C noise a key constraint in high-speed, high-resolution data converters.

The sampled noise follows a Gaussian probability distribution with zero mean and variance σ² = kT/C. Critical properties:

• Each sampling event captures an independent, identically distributed (IID) random value
• The noise is white and uncorrelated from sample to sample
• The probability of a noise sample exceeding ±3σ is only 0.27%
• For a 1 pF capacitor, the standard deviation is approximately 64.4 µV This Gaussian nature means the noise can be averaged down by oversampling, improving SNR by 3 dB per doubling of the sampling rate.

kT/C noise directly limits the achievable Effective Number of Bits (ENOB) in a data converter. For a given full-scale voltage (V_FS) and capacitance:

• Maximum SNR: SNR_max = (V_FS²/2) / (kT/C)
• ENOB limit: ENOB_max = (SNR_dB - 1.76) / 6.02
• Example: A 1 Vpp signal on a 0.5 pF capacitor yields a theoretical maximum SNR of ~73 dB, limiting resolution to approximately 11.8 bits regardless of quantizer precision This establishes a hard physical boundary that distinguishes truly random noise from deterministic quantization error in fingerprinting applications.

Although kT/C noise is fundamentally random, the effective capacitance and operating temperature vary between devices due to:

• Process variations: Oxide thickness and dielectric constant shifts alter actual C values by ±10-15%
• Temperature gradients: Localized heating from adjacent digital circuits raises T, increasing noise power
• Switch non-idealities: Charge injection and clock feedthrough add deterministic offsets to the sampled noise These variations create a device-specific noise pedestal that, while random per sample, exhibits a unique statistical signature in its variance and distribution that can be exploited for hardware authentication.