Brain Simulation Software
BrainX — our in-house open-source brain-simulation ecosystem, spanning ion channels to whole-brain networks, with native heterogeneous-hardware support and a clear path to hundred-billion-neuron whole-brain dynamics.
BrainX: a software ecosystem for whole-brain dynamics simulation
BrainX is our open-source brain-simulation ecosystem, spanning multi-scale modeling from ion channels and single cells to circuits and whole-brain networks. It brings modern AI compilation (JAX/XLA) into computational neuroscience, giving brain-dynamics simulation native support for CPU/GPU/TPU and heterogeneous accelerators, seamless interoperability with the deep-learning ecosystem, and a clear path toward whole-brain dynamics at hundred-billion-neuron scale.
BrainX module matrix
BrainX is composed by layer of responsibility: a runtime and numerics foundation at the bottom, event computing and multi-scale modeling in the middle, with learning algorithms and toolkits above. Each layer is independently usable and designed to compose.
- L01 Runtime State transforms, JIT, cross-device execution
- L02 Numerics Physical units & unit-aware mathematics
- L03 Event Spiking and event-driven sparse operators
- L04 Modeling Cellular, neural-mass and general brain-dynamics modeling
- L05 Learning Online-learning compiler for brain dynamics
- L06 Toolkit General-purpose modeling & analysis tools
Architecture and key capabilities
- JAX-native runtime
- BrainState provides state transforms and JIT compilation, with native CPU/GPU/TPU and heterogeneous-system support and automatic parallel scheduling.
- End-to-end differentiable simulation
- The entire simulation pipeline is differentiable — parameters can be fit to experimental data via gradients, bridging simulation and brain-inspired computing.
- Unit-aware scientific computing
- BrainUnit / SAIUnit embeds physical units inside the high-performance AI computing pipeline, eliminating dimensional errors at the source.
- Large-scale parallel architecture
- BrainEvent sparse operators, BrainMass neural-mass reductions, and distributed scheduling together push toward 10¹¹-neuron simulation.
BrainX powers the algorithm research in the next track and provides the runtime substrate for the whole-brain computational models in Track III. It is the foundation layer of our software–algorithms–models paradigm.