nsram 0.2.0

NS-RAM: Neuro-Synaptic RAM simulator — floating-body transistor neuron/synapse modeling with charge-trapping plasticity

nsram — Neuro-Synaptic RAM Simulator

The first open-source Python library for simulating NS-RAM floating-body transistor neurons with charge-trapping synaptic plasticity.

Based on: Pazos et al., "Synaptic and neural behaviours in a standard silicon transistor", Nature 640, 69-76 (2025).

Installation

pip install nsram              # CPU only (numpy)
pip install nsram[gpu]         # GPU acceleration (PyTorch)
pip install nsram[all]         # Everything (GPU + plotting + ODE solvers)

Quickstart

from nsram import NSRAMReservoir, rc_benchmark

# Create 128-neuron reservoir with heterogeneous short-term plasticity
res = NSRAMReservoir(N=128, stp='heterogeneous')

# Run reservoir computing benchmarks
results = rc_benchmark(res)
# XOR-1: 90.3%, MC: 2.29, NARMA-10: 0.52, Wave-4: 92.5%

What is NS-RAM?

NS-RAM (Neuro-Synaptic Random Access Memory) is a standard CMOS floating-body transistor that exhibits both neuron-like spiking and synapse-like plasticity from device physics alone — no special materials or processes required.

Key device physics simulated:

• Impact ionization avalanche (Chynoweth model): I_aval = I0 × exp((Vcb - BVpar) / Vt)
• Breakdown voltage control: BVpar = 3.5 - 1.5 × Vg1 (gate-tunable, from Pazos SPICE)
• Temperature dependence: BVpar(T) = BVpar × (1 - 21.3μ × ΔT)
• SRH charge trapping: dQ/dt = k_cap(Vg2) × (1-Q) × rate - k_em × Q
• VG2-controlled mode switching (neuron ↔ synapse)

Novel Finding: Charge Trapping = Tsodyks-Markram STP

This library implements the mapping between NS-RAM charge trapping and the Tsodyks-Markram short-term plasticity model:

NS-RAM physics	TM-STP neuroscience
Q (trapped charge)	1 − x (depleted resources)
k_cap(VG2)	U (utilization)
1/k_em	τ_rec (recovery time)
ΔVth = −αQ	PSP amplitude modulation

VG2 voltage controls the STP type: low VG2 → depression, high VG2 → facilitation.

API Reference

NSRAMReservoir(N, stp, connectivity, ...)

Parameter	Default	Description
N	128	Number of neurons
stp	'heterogeneous'	STP mode: 'none', 'std', 'stf', 'heterogeneous'
connectivity	'sparse'	'sparse' (15%), 'small_world', 'dense'
spectral_radius	0.90	Recurrent weight spectral radius
variability	0.10	Die-to-die parameter variability
backend	'auto'	'numpy', 'torch', 'auto'

rc_benchmark(reservoir, n_steps, n_reps)

Runs XOR (τ=1,2,5), Memory Capacity, NARMA-5/10, and 4-class waveform classification.

Lower-level API

from nsram import NSRAMParams, NSRAMNetwork
from nsram.physics import breakdown_voltage, avalanche_current, srh_trapping

# Direct physics calculations
bvpar = breakdown_voltage(Vg1=0.40, temperature=350)
I_aval = avalanche_current(Vcb=2.5, Vg1=0.40)

# Custom network
params = NSRAMParams(N=256, variability=0.20, bg_frac=0.90)
net = NSRAMNetwork(N=256, params=params, connectivity='small_world')
result = net.run(input_signal, noise_sigma=0.01)

Device Parameters

Default parameters from Pazos et al. SPICE model (Zenodo: 13843362):

Parameter	Value	Source
BVpar₀	3.5 V	BJTparams.txt
dBVpar/dVg1	-1.5 V/V	BJTparams.txt
Tbv1	-21.3 μ/K	Davalanche.txt
Is	1×10⁻¹⁶ A	BJTparams.txt
Bf	50	BJTparams.txt
Vth₀ (NMOS)	0.432 V	PTM130bulk_lite.txt
Energy/spike	21 fJ	Nature 640
Cell area (1T)	8 μm²	Nature 640
Cell area (2T)	17 μm²	Nature 640

Citation

If you use this library, please cite:

@article{pazos2025nsram,
  title={Synaptic and neural behaviours in a standard silicon transistor},
  author={Pazos, Sebastian and others},
  journal={Nature},
  volume={640},
  pages={69--76},
  year={2025}
}

License

Apache 2.0