IBM/aihwkit

What it solves

This toolkit addresses the "von Neumann bottleneck," where the separation of memory and computation creates performance limits. It enables the exploration and simulation of in-memory computing, where AI workloads are processed directly within memory devices (like PCM, RRAM, or Flash) to increase speed and energy efficiency.

How it works

The kit provides a PyTorch integration that replaces standard neural network layers with analog versions (fully connected, convolution, LSTM) and introduces analog-aware optimizers. It also includes a high-performance C++ simulator (CUDA-capable) that models the physical characteristics of analog devices, including noise, device-to-device variations, and ADC/DAC discretization.

Who it’s for

Researchers and developers working on hardware-aware AI, neuromorphic computing, and the design of energy-efficient AI accelerators using analog crossbar arrays.

Highlights

• PyTorch Integration: Seamlessly integrate analog neural network modules and training workflows into PyTorch.
• Hardware-Aware Training: Include hardware non-idealities and noise in the forward pass to create more robust models for analog hardware.
• Device Simulation: High-performance simulation of material characteristics and crossbar configurations with adjustable parameters.
• Model Conversion: Utility to automatically convert pre-trained digital models to their analog equivalents.