Simulate morphologically detailed neurons and train networks built from them on GPUs.DeepDendrite turns dendritic dependencies into parallel work without changing the required solve order.
在 GPU 上仿真具有精细形态的神经元，并训练由它们构成的网络。DeepDendrite 将树突依赖关系转化为并行任务，同时保持方程求解所需的运算顺序。
DeepDendrite is built for workloads where morphology, ion-channel dynamics, synaptic location, and spike timing are part of the model—not details to be averaged away.
Run multi-compartment models with heterogeneous mechanisms and explicit dendritic structure, from single cells to populations of morphologically distinct neurons.
Use detailed human pyramidal cells as computational units in a network, propagate teaching signals through dendrites, and evaluate learned responses on image-classification tasks.
Solving cable equations is not an unordered batch operation. During triangularization, children must be processed before their parent; back-substitution reverses that dependency. DHS finds independent compartments that can run together.
Explore the scheduling principle →Convert compartment connectivity into parent–child relationships and calculate node depth.
Select the deepest available candidates and arrange them into dependency-safe parallel levels.
Execute multiple compartments within each neuron and organize data around the GPU memory hierarchy.
The published workflows show how the same framework supports fine-grained biophysical experiments and learning with morphologically detailed neural networks.
Attach 24,994 explicit spines to a reconstructed human pyramidal cell and compare how clustered and distributed synaptic inputs change somatic voltage and spike probability.
Encode images as spike trains, train a 784 → 64 → 10 Human Pyramidal Cell Network, and compare its responses on clean and transfer-adversarial samples.
DeepDendrite combines an optimal tree schedule with GPU-aware storage and a CoreNEURON-based simulation engine. The advantage comes from matching the runtime to the structure of the model.
DHS exposes work across independent branches instead of assigning an entire detailed cell to one serial thread.
Aligned compartment data and register reuse reduce expensive global-memory transactions during the solve.
The GPU runtime supports conventional detailed-neuron simulation alongside the published HPC-Net learning workflow.
Performance depends on the model, hardware, thread allocation, and comparison platform. The benchmark page pairs every headline result with its original paper figure and experimental scope.
View benchmarks and paper figures →Clone the public repository with Git. DeepDendrite currently uses a source-based installation, with GPU and compiler requirements documented in the build guide.
git clone https://github.com/pkuzyc/DeepDendrite.git
cd DeepDendriteHelioX complements the NEURON modeling ecosystem with a GPU-native backend designed around irregular dendritic trees, sparse spikes, continuous-time dynamics, and structure-aware learning.
Model specification remains in the established NEURON workflow. After the model is frozen, HelioX binds its states to a runtime that coordinates simulation, gradient computation, and parameter updates on the GPU.
Morphology, mechanisms, connectivity, and event routing
Runtime identifiers connect model variables to GPU execution
Simulation, analytical gradients, and parameter updates
Move only the event data required when rare spikes occur.
Provide consistent mechanism instantiation and stable variable access.
Use multiple streams while preserving within-step ordering.
Map dependency-safe dendritic work to custom GPU execution.
Bind structure-aware gradients directly to simulator states.
The reported results pair simulation throughput with numerical fidelity, then extend the same runtime to whole-circuit fitting and deep biophysical learning.
Across the four FP64 simulation benchmarks reported in Table 1 on the paper's RTX 5090 setup.
Against NEURON in the reported comparison, with identical spike timing in the tested case.
Compared with 47.4 GB for the CoreNEURON-based baseline on the reported RTX 4090 task.
For the reported three-layer MNIST biophysical MLP comparison in Table 4.
These values describe the paper's tested models, hardware, precision, and baselines; they are not universal performance guarantees.
HelioX: A GPU-Native Framework for Simulation and Training of Biophysically Detailed Networks
This preview summarizes information already disclosed in the paper and presentation. It does not include source code, private APIs, or release details.
View online at ICML ↗DEEPDENDRITE · OPEN-SOURCE RESEARCH SOFTWARE