EDEN: modular neurosimulation with decoupled model description and execution

• The EDEN framework decouples model descriptions from execution for extensible neurosimulation.¹
• The modular paradigm supports computational neuroscience and model-driven BCI research (Frontiers; methods, implementation path on the order of 12–24 months).^{1 1}

Gardner updates

• The EDEN framework decouples model descriptions from execution for extensible neurosimulation (Frontiers), supporting computational neuroscience and model-driven BCI research via a modular model/execution split. ¹

Weekly enrichment (2026-07-20)

• The paper (Panagiotou, Miedema, Soudris, Strydis; Erasmus MC / TU Delft) appeared in Frontiers in Neuroinformatics on Aug 7, 2025 (vol. 19:1572782, DOI 10.3389/fninf.2025.1572782) and reframes simulation engines as interchangeable “backends” behind a common NeuroML model frontend.^{2 3}
• EDEN adopts NeuroML v2 as its input standard so modelers avoid learning a simulator-specific language, while just-in-time code generation targets the available hardware (CPUs, and via backends GPUs, FPGAs, and AI chips).^{2 3}
• Extensibility is demonstrated by integrating two fundamentally different backends: flexHH, an FPGA-based accelerator for extended Hodgkin-Huxley multicompartmental networks, and SpiNNaker, an ARM-based neuromorphic cluster for large-scale spiking networks.^{2 3}
• The flexHH backend was functionally validated against a mathematically equivalent three-compartment inferior-olive neuron model (De Gruijl et al., 2012), then benchmarked on networks of inferior-olive neurons wired as Gilbert G(n, p) random graphs with gap-junction (electrical) coupling.^{2 3}
• flexHH’s fixed execution time per population size makes it advantageous only for large, densely connected networks; it can reportedly handle up to 18,240 neurons at full connectivity on a single FPGA card, whereas EDEN’s generic CPU backend hit a memory limit above 5,760 fully connected neurons (~300 bytes per parsed connection tag, reaching tens of GB for 5,000+ fully connected neurons).^{2 3}
• The SpiNNaker backend was validated on the CUBA leaky-integrate-and-fire benchmark (Brunel, 2000) with n = 4,000 neurons and p = 2% connectivity; inter-spike-interval and coefficient-of-variation distributions matched the CPU backend, with a two-sample Kolmogorov-Smirnov test showing no significant difference (p > 0.1).^{2 3}
• SpiNNaker runs each simulation step in a fixed 1 ms of real time; with the CUBA model’s 0.1 ms time step it executes at 1/10 of real-time speed, so it favors large networks while EDEN’s CPU backend stays competitive at small/medium sizes.^{2 3}
• A planned optimization—aggregating identical linear synapses (Lytton, 1996)—is expected to speed dense point-neuron simulations by at least an order of magnitude; the original 2022 EDEN release already ran NeuroML models one to nearly two orders of magnitude faster than NEURON on a desktop.^{2 4}

Footnotes

1. https://news.google.com/rss/articles/CBMimgFBVV95cUxQeERodkFSRnJROVFYRGo2TGpYQnoyajhkMmc0djEzUFhTWnI2YjVTakVJNGtKWmpkZnJlemMzUXhQZUdDWDFRanZ2UkpIc2ZqLUFSczNIRjdYLVEwUVRoWFdFdG5QTVlOUkdFQkt5UDBfVXhuRUdjZlZ0SHlzVk42Ums4UmFUWFpZM3NMVEMxZ2U0VnJMSzlYYWx3?oc=5 ↩ ↩² ↩³ ↩⁴

2. https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/fninf.2025.1572782/full ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

3. https://pmc.ncbi.nlm.nih.gov/articles/PMC12367680/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

4. https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/fninf.2022.724336/full ↩