analogsim / analoglib

An open-source Python library for simulating Analog In-Memory Computing (IMC) and neural network inference on resistive crossbar architectures (ReRAM, PCM, Flash).

⚡ Key Features

• 🧠 PyTorch & NumPy Importers: Convert PyTorch nn.Module or NumPy weight matrices directly into physical crossbar arrays.
• 📐 Analog Intermediate Representation (AIR): Decouples high-level model definitions from low-level physical crossbar backends.
• 🧱 Tiled Crossbar Architecture (TiledCrossbar): Automatically partitions arbitrary matrix sizes across a 2D grid of physical crossbar tiles with global $w_{\text{max}}$ scale preservation.
• ⚡ Physical Hardware Non-Idealities (analoglib.effects):
  • Parasitic IR Drop: Wordline/bitline wire resistance simulation ($r_{\text{wire}}$ per cell).
  • Thermal Scaling: Arrhenius temperature-dependent conductance shift $G(T) = G_0 \exp(-E_a / k_B T)$.
  • Retention Drift: Power-law temporal decay $G(t) = G_0 (t/t_0)^{-\nu}$.
• 📊 Hardware Profiler & Analytics (AnalogProfiler): Estimates array power (W), read energy (J), ADC/DAC energy overhead, area ($\mu\text{m}^2$), latency (s), and throughput efficiency ($\text{TOPS/W}$).
• 🔌 Circuit Exporter: Export loaded models into ngspice and LTspice netlists for circuit simulation.
• 📦 Encrypted .analog Binary Format: Secure serialization using AES-256-GCM + MsgPack.
• 🛠️ CLI Utility: Command-line tool analog / analogsim for info, simulation, profiling, and SPICE export.

📦 Installation

Install from PyPI

pip install analogsim

Install with Optional Dependencies (PyTorch, Matplotlib)

pip install "analogsim[all]"

Install from Source

git clone https://github.com/Aditya-Raj-Parashar/analoglib.git
cd analoglib
pip install -e .

🚀 Quick Start

1. PyTorch to Analog Simulation Pipeline

import torch.nn as nn
import analogsim as al
import numpy as np

# 1. Define your PyTorch model
torch_model = nn.Sequential(
    nn.Linear(784, 128),
    nn.ReLU(),
    nn.Linear(128, 10),
)

# 2. Convert PyTorch model to AnalogModel via AIR
model = al.AnalogModel.from_torch(torch_model)

# 3. Compile targeting physical ReRAM crossbars + ADC/DAC + Hardware Effects
model.compile(
    device=al.ReRAM(g_min=1e-6, g_max=100e-6, num_states=256, read_noise_sigma=0.01),
    adc_bits=8,
    dac_bits=8,
    r_wire=1.0,    # Parasitic IR drop wire resistance
    E_a=0.1,       # Thermal Arrhenius scaling
    nu=0.05,       # Retention drift exponent
)

# 4. Simulate inference in hardware mode
x_input = np.random.uniform(0, 1, 784)
result = model.simulate(x_input, mode="hardware")

# 5. Generate Hardware Performance & Accuracy Report
result.report()

2. Direct Crossbar Operations

import analogsim as al
import numpy as np

# Define physical ReRAM device
reram = al.ReRAM(g_min=1e-6, g_max=100e-6, num_states=256)

# Create a differential crossbar (128 rows x 64 columns)
xbar = al.Crossbar(128, 64, device=reram, differential=True)

# Load weights (mapped automatically to G+ and G- conductances)
W = np.random.uniform(-1.0, 1.0, (128, 64))
xbar.load_weights(W, quantize=True)

# Perform VMM
V_in = np.random.uniform(0.0, 1.0, 128)
I_out = xbar.vmm(V_in, mode=al.SimulationMode.HARDWARE)

🛠️ CLI Usage

analogsim comes with a command-line tool analog (or analogsim):

# View model metadata & layer details
analog info model.analog

# Run hardware simulation
analog simulate model.analog --mode hardware

# Profile power, latency, area, and TOPS/W
analog profile model.analog

# Export SPICE netlist for ngspice
analog export-spice model.analog --out circuit.cir --dialect ngspice

📜 Publishing to PyPI

To build and publish analogsim to PyPI:

# 1. Install build tools
pip install build twine

# 2. Build source distribution and wheel
python -m build

# 3. Upload to PyPI
python -m twine upload dist/*

📄 License

MIT License © 2026 Aditya Raj Parashar