Nuro — a universal Python SDK for spiking neural networks (train on GPU, deploy to SpiNNaker/Loihi/analog — no code changes)

[Malte Wagenbach]

Hi SpiNNaker community,

We've been building something that directly addresses a pain point we've seen come up repeatedly — the gap between training SNNs and deploying them on neuromorphic hardware.

We're Vantar, and we're building Nuro — a Python SDK that compiles spiking neural networks to any neuromorphic backend. Train with surrogate gradients on GPU (PyTorch-style). Deploy the same network to SpiNNaker, Intel Loihi, or analog neuromorphic chips — with no code changes. One API for the entire neuromorphic ecosystem.

The core idea: you shouldn't have to learn a new toolchain every time you switch hardware, and you shouldn't have to rewrite your network to go from local simulation to real silicon.

A quick example of what working with Nuro looks like:

  import nuro

  net = nuro.Network()
  # define your SNN once
  # train on GPU with surrogate gradients
  net.train(data, backend="gpu")
  # deploy to SpiNNaker — same object, no changes
  net.deploy(backend="spinnaker")

We're in early access now. If you're working on SpiNNaker and want to be one of the first to try it — whether for RL, real-time sensor fusion, or research simulations — you can sign up at vantar.xyz or just reply here.

We'd also love to hear what workflows or pain points matter most to this community. There's clearly a lot of sophisticated work happening here and we want to make sure Nuro is genuinely useful for it.

— The Vantar Team
https://vantar.xyz