aihwkit.inference.noise.reram module

Phenomenological noise models for ReRAM devices for inference.

class aihwkit.inference.noise.reram.ReRamCMONoiseModel(coeff_dict=None, g_max=None, g_min=None, prog_noise_scale=1.0, read_noise_scale=1.0, drift_scale=1.0, decay_dict=None, read_dict=None, acceptance_range=0.02)[source]

Bases: BaseNoiseModel

Noise model inferred from Analog Filamentary Conductive-Metal-Oxide (CMO)/HfOx ReRAM devices from IBM Research Europe - Zurich.

This noise model is estimated from statistical characterization of CMO/HfOx devices from Falcone et al. (In Review)

Programming noise:

Described by a linear function with respect to the G target. Coefficients are considered for two acceptance ranges, 0.2% and 2% of target conductance

Conductance Decay:

Drift in CMO/HfOx devices showed independence of the target conductance value. Mean and STD of the conductance distribution were fitted with 1st-order polynomial as a function of the log(t) where t is the time of inference

TODO: Read noise (1/f) characterization of CMO/HfO<sub>x</sub> available at Lombardo et al. DRC (2024) but not implemented.

Note

To account for short-term read noise (about 1%) one should additional set the forward.w_noise parameter to about 0.01 (with w_noise_type=WeightNoiseType.ADDITIVE_CONSTANT)

Parameters:

  • coeff_dict (Dict[float, List] | None) – acceptance range with coefficients for the programming noise in \(\mu S\),

  • g_converter – Instantiated class of the conductance converter for a single device per

  • cross-point.

  • g_max (float | None) – In \(\mu S\), the maximal conductance, i.e. the value the absolute max of

  • to. (the weights will be mapped)

  • g_min (float | None) – In \(\mu S\), the minimal conductance, i.e. the value the absolute min of

  • to.

  • prog_noise_scale (float) – Scale for the programming noise.

  • read_noise_scale (float) – Scale for the read and accumulated noise.

  • drift_scale (float) – Scale for the drift coefficient.

  • uS. (since the result of the polynomial fit is given in)

  • decay_dict (Dict[str, List] | None) – mean and std coefficients for the drift noise in \(\mu S\),

  • read_dict (Dict[str, float] | None)

  • acceptance_range (float)

apply_drift_noise_to_conductance(g_prog, drift_noise_param, t_inference)[source]

Apply the accumulated noise according to the time of inference.

Will use unique 1st-order polynomial fits the conductance mean shift and standard deviation shift from the ReRAM

Parameters:

  • g_prog (Tensor) – target conductance values that will be used to add noise

  • drift_noise_param (Tensor) – ccoefficients of the mean and std drift

  • t_inference (float) – time of inference. Times in seconds

Returns:

conductances with noise applied

Return type:

Tensor

apply_programming_noise_to_conductance(g_target)[source]

Apply programming noise to a target conductance Tensor.

Programming noise with additive Gaussian noise with conductance dependency of the variance given by a 1st-degree polynomial. Depends of the acceptance range of the program-and-verify loop

Parameters:

g_target (Tensor)

Return type:

Tensor

generate_drift_coefficients(g_target)[source]

Conductance relaxation is independent of the conductance level

Parameters:

g_target (Tensor)

Return type:

Tensor

class aihwkit.inference.noise.reram.ReRamWan2022NoiseModel(coeff_dic=None, g_converter=None, g_max=None, noise_scale=1.0, coeff_g_max_reference=None)[source]

Bases: BaseNoiseModel

Noise model that was inferred from ReRam publication data.

This ReRam model is and approximation to the data published by Wan et al. Nature (2022).

Conductance dependence of the deviations from the target conductance was estimated from the published figures and fitted with a 4-th order polynomial (only 1 sec, 1 day, 2 day).

No separate data is available for read noise (1/f).

Note

To account for short-term read noise (about 1%) one should additional set the forward.w_noise parameter to about 0.01 (with w_noise_type=WeightNoiseType.ADDITIVE_CONSTANT)

Parameters:

  • coeff_dic (Dict[float, List] | None) – polynomial coefficients in \(\mu S\), \(\sum_i c_i \left(\frac{g_t}{g_\max}\right)^i\) for each time. If not given, the fitted measurement is taken at selected time points only

  • g_converter (BaseConductanceConverter | None) – Instantiated class of the conductance converter (defaults to single pair).

  • g_max (float | None) – In \(\mu S\), the maximal conductance, i.e. the value the absolute max of the weights will be mapped to.

  • noise_scale (float) – Additional scale for the noise.

  • coeff_g_max_reference (float | None) – reference \(g_\max\) value when fitting the coefficients, since the result of the polynomial fit is given in uS. If coeff_g_max_reference is not given and coeffs are given explicitely, it will be set to g_max of the conductance converter.

apply_drift_noise_to_conductance(g_prog, g_target, t_inference)[source]

Apply the accumulated noise according to the time of inference.

Will use unique 4th-order polynomial fit to the ReRAM measurements to the target values.

Parameters:

  • g_prog (Tensor) – will be ignored

  • g_target (Tensor) – target conductance values that will be used to add noise

  • t_inference (float) – time of inference.

Returns:

conductances with noise applied

Raises:

ArgumentError – if t_inference is not one of (1, 24*3600, 2*24*3600) seconds (or any user-defined key in coeff_dic), the error will be raised.

Return type:

Tensor

apply_programming_noise_to_conductance(g_target)[source]

Apply programming noise to a target conductance Tensor.

Programming noise with additive Gaussian noise with conductance dependency of the variance given by a 2-degree polynomial.

Parameters:

g_target (Tensor)

Return type:

Tensor

generate_drift_coefficients(g_target)[source]

Return target values as coefficients.

Since ReRAM does not show drift in the usual sense, here simply the target values will given as coefficients to compute the long-term variations on-the-fly

Parameters:

g_target (Tensor)

Return type:

Tensor