Neuromorphic computing trades dense matrix multiplication for sparse, event-driven spikes — networks of artificial neurons that fire only when they need to, on hardware built to exploit that sparsity. Five tools currently anchor the toolchain, from converting an existing PyTorch model down to a spiking form, to simulating custom neuron models, to shipping onto neuromorphic silicon. This is a field guide to what each one is for.
Neuromorphic Intermediate Representation (NIR) is the hardware-agnostic graph format that lets a model move between tools — export once from any NIR-producing tool, deploy to any NIR-compatible simulator or hardware target, the way ONNX decouples a model from any single framework.
Most projects touch two or three of these tools, not all five. The path usually runs from a source model or a from-scratch design, through NIR, and out to a simulator or physical neuromorphic chip.
Convert a trained PyTorch model, or design a spiking architecture from first principles.
NeuroCUDA · snnTorchSerialize the network graph to a hardware-agnostic, NIR format any downstream tool can read.
NIRRun the network at scale, or model custom neurons for neuroscience research, on GPU-generated CUDA code.
GeNNShip to neuromorphic hardware with on-chip learning for a commercial product.
MetaTF → AkidaA PyTorch-to-SNN compiler with GPU, CPU, and Loihi 2 simulator backends — takes a model you've already trained and converts it into a spiking form, ready to target several backends from one artifact.
$ pip install neurocuda
A PyTorch library for training spiking neural networks from scratch, using surrogate-gradient backpropagation through time to get around the non-differentiability of a spike.
A GPU-accelerated SNN simulator that generates optimized CUDA code directly from high-level neuron descriptions, rather than interpreting a generic graph at runtime.
The hardware-agnostic graph format connecting the other four tools — the neuromorphic equivalent of ONNX. Export a model once, and deploy it to any NIR-compatible simulator or hardware target. NeuroCUDA, for instance, exports bit-exact NIR for graphs such as ResNet-18.
A TensorFlow model converter for deployment on BrainChip's Akida neuromorphic processor, with support for on-chip learning rather than a purely fixed, pre-trained model.