Positioning and scope

Positioning and scope#

brainspike is a differentiable point-neuron modeling substrate: the same neurons, synapses, plasticity rules, and projections run NEST-validated biological simulation and train with gradients or online learning.

What sets it apart#

  • Validated like a simulator. The NEST-compatible family keeps NEST’s parameter names and units and is checked against a live NEST install within documented tolerance bands.

  • Differentiable like a training library. Models compile through JAX/XLA and are trainable with surrogate gradients or online learning rules (braintrace).

  • One graph, written contracts. Every network is one Network graph (populations, projections, devices, observers) built from models that declare their ports, states, parameters and spike rules. The same compiled graph is what a simulation runs and what a learner differentiates, so NEST parity validates what users train.

Who it is for#

  1. Computational neuroscientists who want to run NEST models on accelerators, or train them.

  2. Brain-inspired-computing researchers who need biologically constrained spiking networks.

  3. Libraries in the BrainX ecosystem that build on it, such as braintrace for online learning.

Model scope#

  • Point neurons only. Multi-compartment and population/mean-field models are out of scope.

  • No single-unit rate neurons. Rate models (lin_rate, sigmoid_rate, …) are not carried into the new kernel.

Status: kernel redesign in progress#

The package is being moved onto a new kernel, described in the architecture specification (docs/specs/2026-10-21-architecture-review-and-redesign.md). The first milestone (K0) has passed its go/no-go gates; hardening (K1a) is next. While this is in progress:

  • The current 0.1 API is frozen. It keeps working, but receives no new features or fixes; it will be removed when the new kernel replaces it.

  • The new kernel is private (not yet exported from brainspike). It becomes public in a later milestone, together with a migration guide.

  • One front-end. Network will be the only way to build and run a model. The NEST-style Simulator (create / connect / simulate) is kept as a thin layer that builds a Network.

  • Performance. CPU performance of the new kernel has been measured against the current simulator on one benchmark (Brunel network); GPU performance has not been measured yet, and no GPU numbers are claimed.

Out of scope#

  • Detailed multi-compartment morphology.

  • MPI cluster-scale distributed simulation (single-host multi-device execution is in scope).

  • MUSIC / SONATA co-simulation (importing SONATA network descriptions is in scope).

  • e-prop and other online-learning algorithms, which are implemented in braintrace.