Physics Modeling Notes

Physics Modeling Notes

Core mapping

• Input values are mapped to voltage range [V_min, V_max]
• Signed weights are represented with a differential conductance pair
• Current is computed with Ohm's law and summed per column

Differential pair

Signed weight w is represented as:

• G_pos = (w + 1) / 2 * G_max
• G_neg = (1 - w) / 2 * G_max
• Effective current contribution tracks (G_pos - G_neg) * V

Non-idealities

• ADC quantization: current is mapped into discrete levels
• Thermal noise: transient or persistent perturbation on conductances
• Read disturb: neighboring rows drift with repeated reads
• Write endurance: conductance ceiling scales by exp(-lambda * cycles)

Practical tuning

For pretrained or arbitrary weights, tune:

• I_min and I_range to match expected current swing
• G_max to avoid premature clipping
• ADC bit depth for your target quality/speed tradeoff

Physics Modeling Notes

Core mapping

• Input values are mapped to voltage range [V_min, V_max]
• Signed weights are represented with a differential conductance pair
• Current is computed with Ohm's law and summed per column

Differential pair

Signed weight w is represented as:

• G_pos = (w + 1) / 2 * G_max
• G_neg = (1 - w) / 2 * G_max
• Effective current contribution tracks (G_pos - G_neg) * V

Non-idealities

• ADC quantization: current is mapped into discrete levels
• Thermal noise: transient or persistent perturbation on conductances
• Read disturb: neighboring rows drift with repeated reads
• Write endurance: conductance ceiling scales by exp(-lambda * cycles)

Practical tuning

For pretrained or arbitrary weights, tune:

• I_min and I_range to match expected current swing
• G_max to avoid premature clipping
• ADC bit depth for your target quality/speed tradeoff