---
title: "Computational Neuroscience: Simulating Brains"
document_id: sobe-2025-main-ai
document_type: report_section
parent_document: ../report.md
section_id: main-section-13-computational-neuroscience-simulating-brains
section_order: 13
section_count: 17
language: en
license: CC-BY-4.0
---
# Computational Neuroscience: Simulating Brains
> Selective-retrieval section 13 of 17 from the [complete AI-readable report](../report.md). Content is copied without editorial summarization; relative asset, data, and code paths are rebased for this directory.
- **Parent document:** [`report.md`](../report.md)
- **Section ID:** `main-section-13-computational-neuroscience-simulating-brains`
- **Order:** 13 of 17
- **Words:** 9266
- **Previous:** [Connectomics: Brain structure reconstruction](12-connectomics-brain-structure-reconstruction.md)
- **Next:** [Part 3: Appendix](14-part-3-appendix.md)
Building on very early neuronal models like Lapicque's Recherches quantitatives sur l'excitation electrique des nerfs ([Lapicque, 1907](https://link.springer.com/article/10.1007/s00422-007-0189-6)), the quest to formally understand and potentially replicate neural computation gained significant momentum in the mid-20th century. Foundational theoretical work by McCulloch and Pitts ([McCulloch and Pitts,1943](https://link.springer.com/article/10.1007/BF02478259)) proposed a mathematical model of simplified neurons, demonstrating how networks of such units could perform logical computations. Inspired by groundbreaking work such as Edgar Adrian’s on the electrical impulses of neurons (see the Chapter Neural dynamics), and the detailed biophysical modeling of the action potential by Hodgkin and Huxley ([Hodgkin and Huxley, 1952](https://doi.org/10.1113%2Fjphysiol.1952.sp004764)), the idea of engineering brain-like computation emerged. Despite the limitations of the era's technology (vacuum tubes and analog circuits), Marvin Minsky and Dean Edmonds rose to the challenge by constructing the Stochastic Neural Analog Reinforcement Calculator (SNARC) in 1951. This machine, powered by vacuum tubes and motorized potentiometers, was a pioneering hardware system designed to embody a multi-neuron network capable of learning through reinforcement. Though SNARC itself was never formally published and remained relatively obscure compared to the impactful theoretical work of McCulloch and Pitts, the biophysical insights from Hodgkin and Huxley, or later influential developments like Rosenblatt's Perceptron ([Rosenblatt, 1957](https://bpb-us-e2.wpmucdn.com/websites.umass.edu/dist/a/27637/files/2016/03/rosenblatt-1957.pdf)), it nonetheless demonstrated that even a modest array of physical neuron-like units could adapt connections to solve tasks, hinting at the potential later realized in artificial neural networks.
Since then, countless researchers have followed in Minsky and Edmonds’ footsteps, striving to replicate or simulate neural systems in increasingly sophisticated ways. These simulations vary widely in their scope, from the type of simulation (e.g., spiking vs. rate-based models) to the number of neurons and synapses, the hardware and software used, and even the computational cost required to simulate one biological second. However, simulating neurons and synapses is only part of the challenge. For brain models to demonstrate meaningful behavior, they must interact with their environment. This requires embodiment – the ability to encode external stimuli into neural activity and decode that activity into actions. Embodiment does not necessarily demand a physical robotic body; it can occur in a virtual environment, where sensory inputs and motor outputs are simulated alongside the brain model. This process mirrors modern brain-computer interfaces or cochlear implants, where information is encoded and decoded into neural inputs and outputs for specific tasks (e.g., seeing, hearing, memorizing, or speaking). Over time, this interaction can improve iteratively, benefiting from adaptive changes in either the emulated brain or the interface.
To explore these ideas further, we will discuss the computational architectures used to simulate neuronal networks and the tools that enable these digital neurons to interact with digital bodies.
## Methods of Emulating the Brain
Computational models of neurons and synapses have become indispensable tools for bridging the gap between partial functional measurements, structural wiring data, and the richly detailed behavior of living brains. As discussed in prior chapters, exhaustive single-neuron resolution recordings are often unattainable in large organisms, and connectomes alone are insufficient to explain circuit function. Neuron models aim to capture how each cell transforms incoming activity into an output, whether that output is a firing rate, a spike train, or a graded voltage response. By fitting a model’s parameters to observed neuronal behavior, researchers gain explanatory power and a means to predict how real cells might respond under new conditions.
### Neuron Models
Neuron models aim to capture how electrical or synaptic inputs map onto changes in a neuron’s membrane potential or spiking. Some frameworks treat neurons as abstract point units with minimal internal dynamics, whereas others explicitly encode multiple ion channels, compartments, or morphological details. Different modeling choices imply different storage (how many state variables must be updated) and computational costs (how many floating‐point operations per millisecond). Simpler models can be integrated quickly (tens of flops/ms) but may omit key spike features. In contrast, more detailed schemes can require hundreds to thousands of flops/ms while capturing richer physiology. A variety of software frameworks exist to facilitate simulations, including NEURON ([Carnevale and Hines, 2006](https://neuron.yale.edu/neuron/publications/neuron-book)), NEST ([Gewaltig and Diesmann, 2007](http://www.scholarpedia.org/article/NEST_(NEural_Simulation_Tool))), Nengo ([Benkolay et al., 2014](https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/fninf.2013.00048/full)), NeuCube ([Kasabov et al., 2014](https://doi.org/10.1016/j.neunet.2014.01.006)), NeuroGrid ([Benjamin et al., 2014](https://ieeexplore.ieee.org/document/6805187)), GeNN ([Yavuz et al., 2016](https://www.nature.com/articles/srep18854)), Brian2 ([Stimberg et al., 2019](https://pmc.ncbi.nlm.nih.gov/articles/PMC6786860/)), BMTK ([Dai et al., 2020](https://doi.org/10.1371/journal.pcbi.1008386)), and many more.
For a tabular comparison of different models and their respective computational demands, see Figure 2 [Izhikevich, 2004](https://www.izhikevich.org/publications/whichmod.pdf). Due to copyright constraints, we cannot replicate this image here.
The leaky integrate-and-fire (LIF) approach is one of the simplest spiking models. It treats the neuron membrane as a resistor-capacitor circuit, whose membrane potential u integrates an incoming current while spontaneously decaying (the “leak”) towards its resting potential; if the membrane potential exceeds the firing threshold, the neuron spikes. The membrane then resets to a resting voltage. This scheme introduces only one differential equation (plus a rule for thresholding and reset), and captures fundamental aspects of spiking behavior. Its computational lightness makes it a staple for large-scale network simulations, although it omits ion-channel kinetics or dendritic geometry ([Skocik and Long, 2014](https://doi.org/10.1109/TNNLS.2013.2294016)). Because it tracks only one state variable (u) plus a fixed threshold, LIF can simulate a neuron with as few as 4–40 floating‐point operations per millisecond in practice ([Izhikevich, 2004](https://www.izhikevich.org/publications/whichmod.pdf); [Skocik and Long, 2014](https://doi.org/10.1109/TNNLS.2013.2294016)). This low cost, however, means less biophysical realism – one trades away phenomena like bursting or variable spike thresholds for speed and minimal memory usage.
An example of a somewhat more general and potentially more biophysically detailed model is the Izhikevich model, which is a more flexible single-compartment system with two coupled state variables, typically v for membrane voltage and u for a recovery process. Although it lacks direct biophysical details about sodium and potassium currents, it can reproduce many spike-timing patterns – regular, bursting, chattering – through suitable tuning of just four parameters. In practice, Izhikevich neurons are common in spiking network simulations that benefit from more realistic spiking patterns than LIF can provide, but do not require the fully biophysical detail of Hodgkin–Huxley ([Izhikevich, 2004](https://www.izhikevich.org/publications/whichmod.pdf); [Skocik and Long, 2014](https://doi.org/10.1109/TNNLS.2013.2294016)). A caveat is that the model’s properties can shift as the time step changes, so care is needed to ensure consistent solutions at different step sizes ([Skocik and Long, 2014](https://doi.org/10.1109/TNNLS.2013.2294016)). Implementing these two state variables plus four parameters typically requires on the order of 10–40 flops/ms if run coarsely, but can climb to hundreds or thousands of flops/ms for higher accuracy or smaller time steps ([Izhikevich, 2004](https://www.izhikevich.org/publications/whichmod.pdf); [Skocik and Long, 2014](https://doi.org/10.1109/TNNLS.2013.2294016)). In return, modelers gain the ability to represent bursting, rebound spikes, and other complex dynamics that simple LIF neurons cannot capture.
In contrast to these more phenomenological approaches, the Hodgkin–Huxley model preserves explicit ion-channel kinetics derived originally from voltage-clamp experiments on the squid giant axon. Its membrane potential is governed by separate sodium, potassium, and leak conductances, each described by gating variables that follow voltage- and time-dependent transition rates. Although more parameter-intensive, this scheme remains an essential tool for replicating neuronal action potentials' shapes, thresholds, and frequency responses. It can also be adapted to incorporate multiple channel subtypes or altered gating parameters ([Skocik and Long, 2014](https://doi.org/10.1109/TNNLS.2013.2294016)). Beyond single-compartment usage, the Hodgkin–Huxley formalism commonly appears in multi-compartment neuron models, where each compartment has its own channel dynamics. Because it tracks four or more separate state variables and uses multiple exponentials or lookup tables, Hodgkin–Huxley can require anywhere from a few hundred to well over a thousand flops/ms to achieve faithful spike timing. This added cost buys far more physiological detail, enabling accurate reproduction of real spike shapes and voltage‐dependent channel behaviors ([Izhikevich, 2004](https://www.izhikevich.org/publications/whichmod.pdf); [Skocik and Long, 2014](https://doi.org/10.1109/TNNLS.2013.2294016)). Further biological complexity can be introduced by using multiple nonlinear conductances of the Hodgkin-Huxley type modeling a variety of distinct ion channel families with different ionic species (sodium, potassium, calcium, etc.), kinetic properties (fast, slow, transient, persistent, etc.), and gating mechanisms (voltage, ligand, etc.) ([Gouwens et al., 2018](https://doi.org/10.1038/s41467-017-02718-3)).
Expanding Hodgkin–Huxley to a multi-compartment framework allows modeling of dendritic branches, axonal initial segments, and other spatial structures. These models discretize the neuron’s geometry and, via the axial‐current term derived from classical cable theory, couple neighbouring segments so that voltage in one piece can influence the next. Once active conductances are added to each segment, this framework can capture phenomena such as local dendritic Ca²⁺ or NMDA spikes, axonal back-propagation ([Rama et al., 2018](https://doi.org/10.1016/j.conb.2018.02.017)), and region-specific channel gradients. While multi-compartment models certainly can involve dozens or even thousands of compartments ([Herz et al. 2006](https://doi.org/10.1126/science.1127240)), it is thought that in many situations only a small handful, for instance, two to five, are needed to capture soma-dendrite interactions or back-propagating action potentials ([Carlsmith 2020](https://www.openphilanthropy.org/research/how-much-computational-power-does-it-take-to-match-the-human-brain/)). This extra complexity enables phenomena like dendritic coincidence detection, local dendritic plateau potentials, or intricate backpropagation, which can shape the cell’s coding properties in ways that single-compartment models cannot capture. Thus, large compartmental expansions represent a relatively flexible choice: increasing compartments and parameters can boost realism in dendritic computations, at the cost of heavier per-neuron memory usage and computational flops.
At the finest scale, molecular dynamics (MD) simulates individual atoms to understand ion channel mechanisms like permeation and gating ([Roux, 2002](https://www.sciencedirect.com/science/article/pii/S0959440X0200307X); [Alberini et al., 2023](https://pubs.acs.org/doi/10.1021/acs.jctc.2c00990); [Guardiani et al., 2022](https://www.tandfonline.com/doi/full/10.1080/23746149.2022.2080587)). This offers far better biophysical detail but comes at an extreme computational cost. This cost limits typical simulations to short timescales, often just pico- to nanoseconds, extending to only a few microseconds even on high-performance hardware ([Guardiani et al., 2022](https://www.tandfonline.com/doi/full/10.1080/23746149.2022.2080587)). This poses a significant challenge, as critical functional processes like channel gating or ligand binding frequently occur on much longer, millisecond-to-second timescales ([Guardiani et al., 2022](https://www.tandfonline.com/doi/full/10.1080/23746149.2022.2080587)). The extreme computational cost also severely limits the spatial scale feasible for MD simulations. Although there has been progress thanks to both GPU acceleration ([Schoenholz et al, 2019](https://arxiv.org/abs/1912.04232); [Doerr et al, 2021](https://pubs.acs.org/doi/10.1021/acs.jctc.0c01343)) and to dedicated hardware like the Anton 3 supercomputer ([Shaw et al., 2021](https://dl.acm.org/doi/10.1145/3458817.3487397)), it currently seems unrealistic to directly emulate whole brains or even a single neuron at this level of detail for functionally relevant durations. Consequently, prospective emulations will likely not employ MD directly to simulate neural activity. Instead, its primary contribution could be indirect: providing fundamental biophysical data to inform the construction and parameterization of the more computationally efficient, higher-level neuron models required for brain-scale simulation.
The choice of neuron model often reflects a specific project's target level of biological abstraction, as different models capture varying degrees of underlying function. However, perfect replication down to the atomic level at whole-brain scale will plausibly remain infeasible for the foreseeable future. Furthermore, even within the usable range of models, determining the most computationally efficient approach for a given target level of accuracy is complex; different studies reach varying conclusions based on specific benchmarks, accuracy metrics, and implementation choices ([Izhikevich, 2004](https://www.izhikevich.org/publications/whichmod.pdf); [Skocik and Long, 2014](https://doi.org/10.1109/TNNLS.2013.2294016), [Valadez-Godinez et al., 2019](https://doi.org/10.1016/j.neunet.2019.09.026)). Given these uncertainties and the significant cost associated with simulating unnecessary biological detail ([Guardiani et al., 2022](https://doi.org/10.1080/23746149.2022.2080587)), the field must determine empirically what model complexity is necessary and sufficient to produce functionally faithful emulations. Resolving this challenge will not just guide model development but also define data collection and compute resource requirements, ultimately determining the viability of achieving functionally accurate brain emulations for any given level of funding.
### Synapse Models
Synapses transmit signals between neurons, exhibit short-term dynamics based on recent activity, and undergo long-term changes during learning. Computational models of synapses vary from simple rules to detailed biophysical simulations, with different models emphasizing different aspects of synaptic function.
The simplest models treat each synapse as a fixed weight: presynaptic spikes cause instantaneous jumps in postsynaptic current. While computationally efficient, this ignores that real synapses produce responses that rise and decay gradually over time. More realistic models generate a transient response for each spike, using either an alpha function or a difference of exponentials to capture this time course ([Roth and van Rossum, 2009](https://doi.org/10.1007/978-1-4614-7320-6_240-1)). These waveforms can be applied either to conductance (capturing voltage-dependence but requiring more computation) or directly to current. Although these temporal response functions – functions that describe how conductance or current evolves – in their basic form they still abstract away receptor kinetics and potentially ignore voltage dependence. For this reason, a common extension is to include an NMDA component with a voltage-dependent Mg²⁺, typically as a multiplicative factor on the conductance. Going beyond these relatively simple kinetic schemes, more detailed models use kinetic schemes tracking multiple receptor states (closed, open, desensitized), while the most complex simulations also model neurotransmitter diffusion and geometric effects in the synaptic cleft. However, such detailed diffusion models are typically too computationally demanding for network simulations and are more commonly used to study single synapses ([Destexhe et al., 1998](https://papers.cnl.salk.edu/PDFs/Kinetic%20Models%20of%20Synaptic%20Transmission%201998-3229.pdf)).
The models described above capture instantaneous synaptic responses, but real synapses show activity-dependent changes over milliseconds to seconds, exhibiting either depression (decreased effect) or facilitation (increased effect) with repeated activation ([Citri and Malenka, 2007](https://www.nature.com/articles/1301559)). Two main phenomenological models capture these effects: the Tsodyks-Markram model treats synapses as having a pool of resources that deplete with use and recover over time, while Abbott et al's model directly modifies release probability ([Tsodyks et al., 1998](https://doi.org/10.1073/pnas.95.9.5323); [Abbott et al., 1997](https://doi.org/10.1126/science.275.5297.221)). Both models can fit experimental data well despite their simplicity. More mechanistic variants, however, can track multiple vesicle pools, presynaptic calcium, and other factors that drive more complex short-term dynamics.
Whereas short-term changes fade within seconds, synapses also undergo longer-lasting modifications that can persist for hours or days. Such long-term plasticity is thought to underlie learning and memory, and is typically modeled through Hebbian-like rules. Basic Hebbian models strengthen synapses when pre- and postsynaptic neurons are active together, but require stabilizing modifications to prevent runaway growth. Spike-timing-dependent plasticity (STDP) implements this principle at the level of individual spikes: synapses strengthen when presynaptic spikes precede postsynaptic ones by tens of milliseconds and weaken for the reverse order. Still, straightforward STDP rules cannot capture frequency-dependent or burst-dependent effects. More advanced models address these nuances through multi-spike interactions ([Pfister and Gerstner, 2006](https://doi.org/10.1007/s00422-008-0233-1)) or by simulating how calcium influx drives synaptic changes ([Shouval et al., 2002](https://link.springer.com/article/10.1007/s00422-002-0362-x)).
### Model Fitting and Data-Driven Approaches
Regardless of a model’s complexity, its predictive usefulness depends on how effectively its parameters are constrained by experimental measurements ([Almog and Korngreen, 2016](https://doi.org/10.1152/jn.00360.2016)). Historically, most neuron models were constrained by manually tuning a few parameters (e.g., leak conductance, threshold, or channel densities) to match qualitative observations, such as a neuron’s typical firing frequency or spike waveform. Modelers would “hand-fit” the neuronal behavior by trial-and-error until the simulated voltage traces or firing rates closely resembled a reference dataset. Although workable for simple scenarios, this approach often fails to generalize and can obscure parameter degeneracies – different parameter sets may produce similar outputs without revealing which ones are biologically correct.
The effort to more rigorously constrain neuron models has been significantly advanced by detailed electrophysiological recordings, particularly from single neurons in slice preparations. Building on this, early automated optimization strategies emerged that did not require model differentiability. For instance, Druckmann et al. introduced a multi-objective framework using genetic algorithms to fit conductance-based models by comparing multiple electrophysiological features (e.g., spike rate, action potential shape) from simulations to the mean and standard deviation of those features in experimental recordings ([Druckmann et al., 2007](https://doi.org/10.3389/neuro.01.1.1.001.2007)). The application of such stochastic optimization techniques, including genetic algorithms, was made more accessible and standardized by software packages like BluePyOpt ([Van Geit et al., 2016](https://doi.org/10.3389/fninf.2016.00017)), and these genetic algorithm-based approaches enabled the systematic generation of large libraries of biophysically detailed models ([Gouwens et al., 2018](https://www.nature.com/articles/s41467-017-02718-3)). More recently, simulation frameworks supporting automatic differentiation have gained traction, also frequently utilizing such detailed electrophysiological data. When a model’s equations are fully or partially differentiable (as Hodgkin–Huxley usually is), gradient descent, similar to training methods underlying modern AI systems, can be applied to minimize the mismatch between recorded and simulated activity systematically. This has motivated recent work on differentiable simulators that can backpropagate errors from final spike outputs or time-varying voltage traces through an entire model, yielding parameter sets consistent with optical or electrophysiological data ([Deistler et al., 2024](https://doi.org/10.1101/2024.08.21.608979)). In other cases where discontinuities or non-smoothness arise (e.g., certain spiking resets), generative methods can learn to map recordings to model parameters without needing direct backpropagation through each spike event.
For synaptic models, patch-clamp electrophysiology remains the gold standard for parameter fitting. Single-cell recordings characterize basic synaptic transmission parameters like postsynaptic current kinetics and receptor properties, while paired recordings or optogenetic stimulation reveal presynaptic release probability and short-term plasticity dynamics. Studying long-term plasticity poses additional challenges, as these changes unfold over hours to days and involve complex molecular cascades that are difficult to monitor in living tissue while maintaining sufficient temporal and spatial resolution.
While such approaches benefit from biologically informed model architectures (e.g., known ion-channel structure or a leaky-integrate-and-fire scheme), one can also relax these biophysical priors entirely and rely on purely data-driven methods. Large recurrent or feedforward neural networks – trained end-to-end on high-dimensional time-series data – can, in principle, capture dynamics that conventional equations might omit ([Wang et al., 2025](https://doi.org/10.1038/s41586-025-08829-y)). However, this freedom demands enormous datasets, with recent scaling analyses suggesting that even for C. elegans, almost an order of magnitude more neural recordings than currently exist would be required for highly predictive models ([Simeon et al., 2024](https://doi.org/10.1101/2024.02.13.580186)). Nonetheless, in smaller, relatively stereotyped organisms where single-neuron resolution data could be collected in large quantities, purely “black-box” models with minimal biophysical assumptions could at least theoretically thrive.
### Structure to function
For larger mammalian brains, though, comprehensive, whole-brain functional data remain beyond present technology, making it impractical to learn all parameters directly from raw observations. Consequently, biophysical priors – such as those embedded in multi-compartment Hodgkin–Huxley or simpler spiking models – remain crucial to reduce parameter space and keep simulations biologically plausible. Yet even these more traditional approaches struggle to scale up to billions of cells; far more extensive measurement capabilities would be required to fit every neuron and synapse.
A promising strategy for bridging the structure-to-function gap is to develop generative models that link morphological and molecular data to neuronal function. In smaller, more experimentally tractable organisms such as C. elegans or larval zebrafish, one can gather both a molecularly annotated connectome and rich functional data (e.g., through optogenetic perturbations and whole-brain recordings). These models learn how specific structural attributes translate into core biophysical parameters like channel conductances or short-term plasticity kinetics by correlating neuronal responses with ultrastructural features, such as synapse size, receptor distributions, and dendritic geometry. Once validated in systems with abundant ground-truth data, the same approach can be extended to large mammalian brains, where whole-brain functional measurements remain infeasible. In this context, Holler et al stands as a landmark demonstration of how morphological information alone can inform functional predictions: by combining slice electrophysiology with electron microscopy, they showed that synapse size reliably predicts synaptic strength in mouse somatosensory cortex ([Holler et al., 2021](https://doi.org/10.1038/s41586-020-03134-2)). Their work exemplifies how structural features can serve as powerful proxies for synaptic function. Incorporating molecular-level data into such pipelines would enable generative frameworks to infer parameters with even greater precision and scale, providing a viable path toward biologically grounded simulations of large neural circuits. The necessary condition for that is an extensive database of slice synaptic electrophysiology data, such as [Campagnola et al, 2022](https://doi.org/10.1126/science.abj5861). In this way, integrating high-resolution structural data with targeted functional measurements can progressively narrow – and ultimately bridge – the gap between brain structure and brain function. This will require substantial investments in such aligned datasets.
### Embodiment
Brains do not function in isolation. They exist within bodies that provide sensory input and execute motor commands, creating closed sensorimotor loops that drive behavior and ground identity. This interdependence raises several questions for whole brain emulation. Firstly, is embodiment (providing the emulation with sensory inputs and motor outputs linked to an environment) a prerequisite for faithful emulation? Secondly, if embodiment is necessary, what level of fidelity is required? Is a generic, functionally adequate body sufficient, or must the emulation replicate the original body's specific physical characteristics and biochemical properties to preserve personal identity and memories?
Arguments supporting the necessity of embodiment, potentially at high fidelity, often point first to the consequences of disrupting brain-body interaction. Functional deficits following sensory or motor deprivation, for instance, are frequently cited to underscore the requirement for continuous interaction simply to preserve long-term stable brain function ([Sandberg and Bostrom, 2008](https://ora.ox.ac.uk/objects/uuid:a6880196-34c7-47a0-80f1-74d32ab98788)). Beyond this need for basic interaction, the case for replicating the specific original body posits that crucial aspects of identity are deeply intertwined with individual physical characteristics. From this perspective, learned motor skills and reflexes are often precisely tuned to the unique biomechanics of the original body (such as limb dimensions, mass distribution, and muscle properties) through adaptations occurring in both the brain and spinal cord ([Shadmehr and Mussa-Ivaldi, 1994](https://doi.org/10.1523/JNEUROSCI.14-05-03208.1994)). The concern raised by this view is that transferring an emulation to a generic or different body might invalidate these learned abilities ([Sandberg and Bostrom, 2008](https://ora.ox.ac.uk/objects/uuid:a6880196-34c7-47a0-80f1-74d32ab98788)). This viewpoint further highlights that the brain's ongoing state and subjective experience are constantly modulated by the body's hormones, metabolites, and biochemistry more generally. The implication suggested is that accurately capturing an individual's baseline temperament, drives, and even subjectively essential aspects of identity requires simulating this precise internal chemical environment ([Sandberg and Bostrom, 2008](https://www.fhi.ox.ac.uk/brain-emulation-roadmap-report.pdf); [McKenzie, 2022](https://brainpreservation.github.io/Embodied_Brain)). Thus, the conclusion drawn from this rationale is that while some form of embodiment might address basic functional viability, a faithful emulation could require reconstructing the original body with sufficient accuracy ([Sandberg and Bostrom, 2008](https://ora.ox.ac.uk/objects/uuid:a6880196-34c7-47a0-80f1-74d32ab98788)).
Counterarguments emphasize the centrality of the brain for identity. This perspective often distinguishes between enduring personal identity, rooted in long-term memories and personality traits presumed encoded within the brain's structure and dynamics, and transient physiological states, such as hormonal influences or mood fluctuations, primarily mediated by the rest of the body ([McKenzie, 2022](https://brainpreservation.github.io/Embodied_Brain)). This view supports the observation that significant body portions, including limbs, organs, and even sensory apparatus (such as cochleas or retinas), can be functionally replaced or lost without fundamentally erasing personal identity or core memories ([McKenzie, 2022](https://brainpreservation.github.io/Embodied_Brain)). Further, the topographic mapping of sensory and motor functions within the brain provides a plausible pathway for interfacing an emulation with sufficiently structured sensory inputs and motor outputs without the original peripheral structures. Thus, this position suggests that an 'adequate' body simulation needs only to provide a functional feedback loop with the environment, provided that the brain emulation can adapt (for example, through emulated neuroplasticity).
Encouragingly, the development of the simulation tools needed to investigate these questions and provide embodiment for WBE is advancing. Dedicated open-source platforms like OpenSim from the biomechanics community offer increasingly sophisticated capabilities for detailed musculoskeletal modeling and analysis ([Seth et al., 2018](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1006223)). Concurrently, the field is increasingly leveraging progress in high-performance physics engines such as MuJoCo ([Todorov et al., 2012](https://ieeexplore.ieee.org/document/6386109); [Vaxenberg et al., 2024](https://www.biorxiv.org/content/10.1101/2024.03.11.584515v2.full); [Wang-Chen et al., 2024](https://www.biorxiv.org/content/10.1101/2023.09.18.556649v4)). Initially driven by robotics and reinforcement learning, these engines prioritize simulation speed and robust contact handling, while offering features like GPU acceleration and differentiability, both highly valuable for complex embodied systems. Although significant challenges persist, continued progress in these areas holds the promise of enabling researchers to finally settle these philosophical debates surrounding embodiment empirically.
### Verification
The history of technological advancement is often marked by the establishment of benchmarks. In artificial intelligence, the introduction of ImageNet ([Deng et al., 2009](https://doi.org/10.1109/CVPR.2009.5206848)) and its associated ILSVRC competition revolutionized the field by providing a standardized dataset and objective metrics for comparing models. While these benchmarks did not settle debates about the nature of “intelligence,” they enabled measurable progress by focusing on a tangible proxy: classification accuracy. Similarly, the pursuit of emulating biological brains requires rigorous methods to evaluate success. The field risks stagnation without agreed-upon benchmarks, with progress hindered by subjective or incompatible criteria. Standardized evaluations – even imperfect ones – are crucial for aligning efforts, tracking advancements, and accelerating discovery, much as ImageNet catalyzed the rise of deep learning. The central challenge, therefore, is to design benchmarks that capture measurable aspects of neural or functional fidelity, driving progress until their limitations necessitate the next leap forward.
A foundational framework for evaluation was proposed in Sandberg and Bostrom’s Whole Brain Emulation Roadmap ([Sandberg and Bostrom, 2008](https://ora.ox.ac.uk/objects/uuid:a6880196-34c7-47a0-80f1-74d32ab98788)). They distinguished between simulations, which replicate outputs, and emulations, which replicate internal dynamics, arguing that valid models must preserve the brain’s causal structure – the step-by-step relationships between neural states. However, practical challenges abound. Even setting aside the brain’s potential chaotic dynamics, minor discrepancies – stemming from noise in neural recordings ([Rupprecht, 2021](https://gcamp6f.com/2021/10/04/large-scale-calcium-imaging-noise-levels/)) or parameter inaccuracies – can compound over time, causing simulations to diverge from biological trajectories. To address this, Sandberg and Bostrom suggested tolerating deviations smaller than the brain’s inherent “noise floor,” the variability observed across repeated biological trials. This pragmatic approach underpins one of the field’s most advanced benchmarks, ZAPBench ([Lueckmann et al., 2024](https://openreview.net/forum?id=oCHsDpyawq)), which evaluates larval zebrafish emulations by predicting 30 seconds of neural activity from 10 seconds of observed calcium traces (sampled at 1 Hz). Models are scored on per-neuron mean absolute error (MAE) across over 70,000 neurons.
An alternative framework shifts the focus from neural activity to behavioral indistinguishability. Zador et al. propose an “embodied Turing test”: if a virtual animal navigates novel environments, responds to threats, or learns tasks indistinguishably from its biological counterpart, the emulation succeeds – regardless of internal mismatches ([Zador et al., 2023](https://www.nature.com/articles/s41467-023-37180-x)). This mirrors Turing’s original vision, replacing conversation with sensorimotor behavior. For instance, an artificial beaver might be tested on dam-building, or a simulated fruit fly on evasive maneuvers during flight. However, this framework remains comparatively underdeveloped, existing more as a conceptual proposal than a fully fleshed-out benchmark.
The lack of consensus in brain emulation mirrors debates in AI. Just as no single metric – ImageNet accuracy, chess Elo ratings, or language modeling capability – fully captures “intelligence,” no single benchmark will resolve what defines a successful brain emulation. A promising approach, then, is to develop benchmark suites that aggregate diverse tasks. Initiatives like BIG-Bench ([Srivastava et al., 2022](https://doi.org/10.48550/arXiv.2206.04615)) and HELM ([Liang et al., 2022](https://doi.org/10.48550/arXiv.2211.09110)) evaluate language models across hundreds of scenarios, recognizing that no individual task tells the whole story. A parallel and influential effort in computational neuroscience is Brain-Score.org ([Schrimpf et al., 2020](https://doi.org/10.1016/j.neuron.2020.07.040)), which provides an integrative platform for evaluating how well computational models, particularly those of the visual system, align with a diverse array of neural and behavioral benchmarks. By scoring models on their ability to predict brain activity (e.g., in visual cortical areas) and match primate behavioral performance, Brain-Score aims to drive progress towards more neurally mechanistic models of specific intelligent functions. For whole brain emulation, a similar suite might combine neural activity prediction (as in ZAPBench), embodied behavior tests, causal perturbation experiments (e.g., optogenetic interventions), and new innovative metrics that the community will develop. Indeed, while benchmark suites and, in general, the portfolio approach to evaluation discussed here provide a more comprehensive evaluation framework, significant work remains to refine and expand the evaluation landscape. Nonetheless, by iteratively improving and expanding benchmarks, the field can advance objectively, grounded in empirical progress rather than abstract debates.
As part of this report, simulation attempts discussed for different organisms were rated on the following 0-3 point scale across 10 dimensions—the dimensions gesture at the breadth of sophisticated benchmarks that eventually need to be developed.
------------------------------------------------------------------------
## Hardware requirements
Before running a brain emulation, models often need to be fitted or trained to match experimental data, which can be computationally intensive. The resources required for this fitting phase depend heavily on numerous factors, including the complexity of the model being fitted (e.g., number of parameters, compartments), the volume and type of data used as constraints, and the specific optimization algorithms employed. Due to this high degree of variability and a lack of standardized benchmarks or comprehensive references quantifying these costs across different scenarios, providing reliable general estimates is currently challenging. Therefore, although model fitting is a critical and potentially resource-intensive step, this report will focus on the computational demands of running brain emulations, rather than fitting them.
Turning to the demands of running brain emulations, hardware requirements remain radically uncertain, primarily because we don't yet know what level of biological detail is necessary for faithful emulation. The Whole Brain Emulation Roadmap ([Sandberg and Bostrom, 2008](https://ora.ox.ac.uk/objects/uuid:a6880196-34c7-47a0-80f1-74d32ab98788)) highlighted this uncertainty, with computational demands for a human brain ranging from 10¹⁵ to 10³⁰ FLOPS: a 15-order-of-magnitude difference. Storage requirements similarly varied from 5×10¹⁵ to 1.12×10²⁸ bits. Precise resource requirements will remain elusive until a consensus emerges on what constitutes a sufficient level of biophysical or functional fidelity.
What is certain is the dramatic advancement in computational power since the roadmap's publication in 2008. Total compute in supercomputers has been rising exponentially for decades, and supercomputers in the 2020s have been reaching exaflop territory, e.g., with El Capitan at more than 1.5 exaflops ([Thomas, 2024](https://www.llnl.gov/article/52061/lawrence-livermore-national-laboratorys-el-capitan-verified-worlds-fastest-supercomputer)), with typical supercomputer costs of \$600M ([Joseph, 2023](https://hyperionresearch.com/wp-content/uploads/2023/11/Hyperion-Research-SC23-Briefing-Novermber-2023_Combined.pdf)). Further, it’s worth noting that clusters used for machine learning applications typically do not appear in TOP500 and that, as a result, the computational capacity of state-of-the-art systems likely exceeds these values. For example, in 2023 Google demonstrated the ability to train language models on a cluster of 50,944 TPUv5e chips (for a peak performance of over 10 FP16 exaFLOPs) ([Ananthamaran, 2023](https://cloud.google.com/blog/products/compute/the-worlds-largest-distributed-llm-training-job-on-tpu-v5e)), and although no public benchmarks have been reported so far, the 100K H100 Colossus cluster recently assembled by xAI would at least theoretically be capable of over 100 FP16 exaFLOPs, with further plans to roughly double the size of the cluster in the coming months ([Mantel, 2024](https://www.heise.de/en/news/xAI-has-apparently-completed-the-world-s-fastest-supercomputer-9857540.html)).
What kind of emulations could such systems run? To provide concrete, though highly simplified, reference points, we analyze two scenarios: a "simplified lower-bound" scenario, characterized by leaky integrate-and-fire (LIF) neurons and alpha-function synapses, aiming to capture minimal spiking behavior, and an "illustrative moderately complex" scenario, characterized by five-compartment Hodgkin-Huxley (HH) neurons and Tsodyks-Markram synapses, likely capable of capturing dendritic computation ([Carlsmith, 2020](https://www.openphilanthropy.org/research/how-much-computational-power-does-it-take-to-match-the-human-brain/)) and short-term synaptic plasticity, respectively. It is crucial to emphasize that our "moderately complex" scenario is still far from the level of detail that might ultimately be required for full biophysical realism; for instance, highly detailed models could necessitate hundreds of compartments per neuron, or potentially even more complex representations whose requirements are currently an open research question (Skuhersky, 2024, personal communication). Such per-neuron or per-synapse complexity increases would lead to proportionally larger storage and computational demands than estimated here. Furthermore, it's important to note that neither of our defined scenarios represents a true upper bound on computational requirements, as both omit other potentially crucial mechanisms like neuromodulation and glial interactions. They also do not involve long-term plasticity mechanisms like STDP and LTP, leading to emulations that are not capable of forming new memories. Nonetheless, these reference points bracket many modeling efforts today and can offer some insight into the computational requirements of emulations involving different organisms. We follow Skocik and Long’s operation-counting approach ([Skocik and Long, 2014](https://doi.org/10.1109/TNNLS.2013.2294016)), and extrapolate to whole brain compute requirements (using whole-brain estimates for mouse and human, though obtaining reliable whole-brain synapse counts can generally be challenging) – compute requirements based on known or estimated neuron and synapse counts (leaving larval zebrafish aside, for which we could not identify reliable synapse count estimates).
As part of these simplified estimates for time-based emulation compute requirements, assuming FP16 operations, we find that both C. elegans and the fly brain are within reach of modern GPUs, that the mouse brain would likely require a relatively small multi-GPU cluster, and that emulating the human brain would require a large-scale cluster – indeed, Lu et al. recently used 14,000 GPUs for their human-scale model ([Lu et al., 2024](https://www.nature.com/articles/s43588-024-00731-3)) – though not beyond the capabilities of frontier systems like xAI’s Colossus. It’s important to emphasize, however, the limitations of these estimates. These FLOPS estimates represent the theoretical computational load of the core model equations. Actual performance on hardware will also depend on factors such as software implementation efficiency, memory bandwidth limitations, interconnect speeds, and hardware utilization, which can lead to practical simulation times differing from what raw FLOPS might suggest. Thus, these estimates neglect other potentially important mechanisms by focusing solely on neuron and synapse models. Further, even within this modeling paradigm, it’s worth noting that the literature on computational requirements for synapse models is far less extensive than for neuron models, despite synapses often dominating computational costs. Finally, these estimates do not consider numerous other relevant factors in actual modeling efforts, including the specific simulation software used, the achieved hardware utilization rate, and more.
Figure 27 - Estimated Compute Requirements for Time-Based Brain Emulation. The computational power (FP16 FLOPS) required for time-based whole-brain emulations across four organisms (orange) compared with the peak performance of reference hardware (grey). The y-axis is on a logarithmic scale. Error bars reflect the range of requirements based on varying assumptions about the underlying biophysical model complexity (e.g., from simple point neurons to multi-compartment models). Calculations and underlying data for this figure are available in the [linked data repository](https://docs.google.com/spreadsheets/d/1so1BojMiuSTadc6BvGAkQL-CM9ieRaUkmFYWh4iGmYc/edit?gid=162927166#gid=162927166).
### Machine-readable figure record: `main-fig-27`
```yaml
id: main-fig-27
document_id: sobe-2025-main-ai
document_path: report.md
anchor: main-fig-27
canonical_url: https://brainemulation.mxschons.com/ai/report.md#main-fig-27
figure: "Time-based emulation compute"
assets:
- "../assets/report/main-fig-27-publication.png"
data_status: final_circulation_export_raster_with_contextual_current_sources
generator: ../code/run_all_figures.py
generator_function: generate_emulation_requirements
publication_export_members:
- "images/image2.png"
canonical_pdf_pages:
- 157
```
#### Inline contextual data (not an exact publication input): hardcoded estimates
```tsv
organism mid_fp16_flops error_fp16_flops
C. elegans 1.87e9 8.25e8
Fly 3.76e12 1.05e12
Mouse 8.88e15 2.13e15
Human 1.12e19 2.70e18
reference:H100 1e15 0
reference:xAI_Colossus 1e20 0
```
#### Figure-generation source map (valid Python; not standalone plotting code)
```python
GENERATOR_PATH = "../code/run_all_figures.py"
SOURCE_FUNCTION = 'generate_emulation_requirements'
SELECTED_SOURCE_LINES = (
'compute_mids = np.array([1.87e9, 3.76e12, 8.88e15, 1.12e19])',
'compute_errs = np.array([8.25e8, 1.05e12, 2.13e15, 2.7e18])',
'ax.bar(x_pos_org, compute_mids, yerr=compute_errs, alpha=0.85, color=GOLD,',
)
```
Indeed, one reason why it can be hard to achieve high hardware utilization rates is storage requirements. The storage requirements estimated here refer to the memory needed to hold the dynamic state variables of the neurons and synapses during active simulation, typically residing in system RAM or the High Bandwidth Memory (HBM) of accelerators like GPUs. Storage requirements can exceed the high bandwidth memory of the system, and even when memory itself is sufficient, the interconnect between different chips can be a bottleneck. Over the past 30 years, the rate of improvement in processing power has far exceeded the rate of improvement in memory and interconnect development. Peak hardware FLOPS have improved by roughly 3x every 2 years, compared to only 1.6x and 1.4x for memory and interconnect bandwidth, respectively, leading to the so-called memory wall ([Gholami et al., 2024](https://arxiv.org/abs/2403.14123v1); [An et al., 2024](https://arxiv.org/abs/2408.14158)). Memory capacity has also undergone a similar trend: today’s El Capitan boasts 1.74 exaFLOPS and over 5.4 petabytes of HBM3 memory ([Thomas, 2024](https://www.llnl.gov/article/52061/lawrence-livermore-national-laboratorys-el-capitan-verified-worlds-fastest-supercomputer)); NEC’s Earth Simulator boasted 41 teraFLOPS and 10 terabytes of DRAM memory in 2004 ([Sato, 2004](https://www.sciencedirect.com/science/article/abs/pii/S0920563203025118?)); representing an approximately 42,400-fold increase in processing power and just a 540-fold increase in memory capacity. Like many other workloads, computational modeling of brain tissue has been affected by this trend: a 2014 study identified interconnect bandwidth as key bottleneck for spiking neural network simulations ([Kunkel et al., 2014](https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/fninf.2014.00078/full)) and a recent in-depth analysis found memory bandwidth and interconnect latency to represent key bottlenecks for all types of neuron simulations studied, from point neurons with current-based synapses to multicompartmental models with conductance-based synapses, particularly at high fan-ins and neuron counts ([Cremonesi et al., 2020](https://journals.sagepub.com/doi/full/10.1177/1094342020912528)).
Figure 28 - Divergent Growth Rates of Compute, Memory, and Interconnect Performance. Exponential growth trends for hardware performance over the past two decades, plotting peak computational throughput (FLOPS, black), memory bandwidth (DRAM BW, green), and interconnect bandwidth (Interconnect BW, purple). The y-axis is logarithmic, with performance values normalized to the peak FLOPS of the [R10000](https://en.wikipedia.org/wiki/R10000) processor. The steeper slope for FLOPS (3.0x improvement every two years) compared to memory (1.6x) and interconnect (1.4x) bandwidth illustrates the growing gap between processing power and data access speeds, a phenomenon known as the "memory wall."
### Machine-readable figure record: `main-fig-28`
```yaml
id: main-fig-28
document_id: sobe-2025-main-ai
document_path: report.md
anchor: main-fig-28
canonical_url: https://brainemulation.mxschons.com/ai/report.md#main-fig-28
figure: "Hardware scaling"
assets:
- "../assets/report/main-fig-28-publication.png"
data_status: final_circulation_export_raster_with_contextual_current_sources
data_sources:
- "../data/legacy/hardware-scaling-flops-bandwidth.csv"
generator: ../code/run_all_figures.py
generator_function: generate_hardware_scaling
publication_export_members:
- "images/image50.png"
canonical_pdf_pages:
- 158
```
#### Inline contextual data (not an exact publication input): hardware series
```tsv
category name year value
HW FLOPS Pentium 4 2005 90
HW FLOPS GTX 580 2010 6000
HW FLOPS K40 2014 15000
HW FLOPS KNL 2016 22000
HW FLOPS TPUv3 2018 420000
HW FLOPS A100 2020 1250000
HW FLOPS H100 2022 2000000
HW FLOPS Gaudi 2 2022 1100000
HW FLOPS B200 2024 4500000
DRAM BW GDDR4 2007 1.8
DRAM BW GDDR5 2010 10
DRAM BW HBM 2015 30
DRAM BW HBM 2 2018 70
DRAM BW HBM 3 2021 120
DRAM BW HBM3E 2024 250
Interconnect BW PCIe 2.0 2007 1.2
Interconnect BW PCIe 3.0 2011 2
Interconnect BW NVLink 1.0 2016 3
Interconnect BW PCIe 5.0 2019 8
Interconnect BW NVLink 4.0 2024 50
```
#### Figure-generation source map (valid Python; not standalone plotting code)
```python
GENERATOR_PATH = "../code/run_all_figures.py"
SOURCE_FUNCTION = 'generate_hardware_scaling'
SELECTED_SOURCE_LINES = (
'dram_bw = {',
'interconnect_bw = {',
"ax.scatter(dram_bw['year'], dram_bw['value'], color=color_dram, s=80, zorder=5)",
"for i, name in enumerate(dram_bw['name']):",
"ax.annotate(name, (dram_bw['year'][i], dram_bw['value'][i] * 1.3),",
"ax.scatter(interconnect_bw['year'], interconnect_bw['value'], color=color_interconnect, s=80, zorder=5)",
"for i, name in enumerate(interconnect_bw['name']):",
"ax.annotate(name, (interconnect_bw['year'][i], interconnect_bw['value'][i] * 1.3),",
"slope_dram, intercept_dram = fit_trend(dram_bw['year'], dram_bw['value'])",
"slope_inter, intercept_inter = fit_trend(interconnect_bw['year'], interconnect_bw['value'])",
'flops_20yr = 10 ** (slope_flops * 20)',
'dram_20yr = 10 ** (slope_dram * 20)',
'flops_2yr = 10 ** (slope_flops * 2)',
'dram_2yr = 10 ** (slope_dram * 2)',
"ax.text(0.18, 0.98, f'{flops_20yr:.0f}x / 20 yrs ({flops_2yr:.1f}x/2yrs)', transform=ax.transAxes,",
"ax.text(0.18, 0.93, f'{dram_20yr:.0f}x / 20 yrs ({dram_2yr:.1f}x/2yrs)', transform=ax.transAxes,",
)
```
Similarly to the compute requirements discussed above, our simplified estimates suggest that, assuming 32-bit state variables, C. elegans and Drosophila brain emulations would fit comfortably within a modern GPU’s available HBM. For an emulation of a mouse brain in our "illustrative moderately complex" scenario (5-compartment HH), it would likely require a somewhat small cluster, and an emulation of the human brain in the same scenario would require a large-scale cluster close in capability to ones such as xAI’s Colossus. The precise storage per element for more complex models will vary based on the specific ion channels and state variables included. These estimates, however, make the important assumption that interconnects do not become the bottleneck, an assumption that may not hold in practice. Data movement and latency bottlenecks are now the primary constraint in training frontier AI systems, the main application of today's largest GPU/TPU clusters ([Erdil and Schneider-Joseph, 2024](https://arxiv.org/abs/2411.01137)). It's also worth noting the stark contrast in software maturity: considerable effort has gone into developing strategies for efficient distributed training of large-scale AI systems over the past decade, while computational neuroscience has only recently begun to leverage GPUs and TPUs at scale. Indeed, while a few pioneering efforts involved large-scale GPU clusters, most computational neuroscience research remains focused on single-GPU or small-cluster implementations where interconnect bottlenecks are less prominent.
Figure 29 - Estimated Memory Requirements for Brain Emulation. The runtime memory (Bytes) required to store the state of whole-brain emulations across four organisms (light blue) compared with the memory capacity of reference hardware (grey). The y-axis is on a logarithmic scale. Error bars reflect the range of requirements based on the same varying assumptions about biophysical model complexity used for compute estimates. Calculations and underlying data for this figure are available in the [linked data repository](https://docs.google.com/spreadsheets/d/1so1BojMiuSTadc6BvGAkQL-CM9ieRaUkmFYWh4iGmYc/edit?gid=162927166#gid=162927166).
### Machine-readable figure record: `main-fig-29`
```yaml
id: main-fig-29
document_id: sobe-2025-main-ai
document_path: report.md
anchor: main-fig-29
canonical_url: https://brainemulation.mxschons.com/ai/report.md#main-fig-29
figure: "Brain emulation storage"
assets:
- "../assets/report/main-fig-29-publication.png"
data_status: final_circulation_export_raster_with_contextual_current_sources
generator: ../code/run_all_figures.py
generator_function: generate_emulation_requirements
publication_export_members:
- "images/image5.png"
canonical_pdf_pages:
- 159
```
#### Inline contextual data (not an exact publication input): hardcoded estimates
```tsv
organism mid_bytes error_bytes
C. elegans 2.60e5 9.40e4
Fly 6.54e8 2.21e8
Mouse 1.63e12 5.45e11
Human 2.05e15 6.85e14
reference:H100 8e10 0
reference:xAI_Colossus 1e16 0
```
#### Figure-generation source map (valid Python; not standalone plotting code)
```python
GENERATOR_PATH = "../code/run_all_figures.py"
SOURCE_FUNCTION = 'generate_emulation_requirements'
SELECTED_SOURCE_LINES = (
'storage_mids = np.array([2.6e5, 6.54e8, 1.63e12, 2.05e15])',
'storage_errs = np.array([9.4e4, 2.21e8, 5.45e11, 6.85e14])',
'ref_storage = [np.nan, np.nan, np.nan, np.nan, 8E+10, 1E+16]',
'ax.bar(x_pos_org, storage_mids, yerr=storage_errs, alpha=0.85, color=TEAL,',
"ax.bar(x_pos_full, ref_storage, alpha=0.7, color=COLORS['caption'],",
"save_figure(fig, 'brain-emulation-storage-requirements')",
)
```
Interconnects can become the bottleneck partly because time-based simulations require constant updates for all neurons and synapses, typically every 0.1 milliseconds. An event-driven modeling paradigm can lower bandwidth demands by instead only requiring such computations when spikes and synaptic events happen, thus also decreasing compute requirements by a couple orders of magnitude. This is typical of neuromorphic systems like IBM's TrueNorth in 2014 (implementing 1 million neurons and 256 million synapses per chip) ([Akopyan et al., 2015](https://ieeexplore.ieee.org/document/7229264)), Intel's Loihi platforms ([Davies et al., 2018](https://ieeexplore.ieee.org/document/8259423), [Orchard et al., 2021](https://ieeexplore.ieee.org/abstract/document/9605018?casa_token=MFuoGbjVeWcAAAAA:HHlmz4QFBgBbMhIY31daj3uPMg_XqjwLkxmvEjmpRoxDtbH_paO3f6t38NUZD_6tAzlX1Q)) and more recent developments like the Hala Point system at Sandia National Laboratories, with over 1.15 billion neurons and 128 billion synapses ([Intel, 2024](https://www.intel.com/content/www/us/en/newsroom/news/intel-builds-worlds-largest-neuromorphic-system.html)). Support for some event-driven operations is also present in some GPU/TPU-based frameworks like BrainPy ([Wang et al., 2023](https://doi.org/10.7554/eLife.86365)), though this is still far from common. Should event-driven modeling be more widely adopted, not every model would necessarily see significant benefits. Indeed, some, like Hodgkin-Huxley neurons, can be difficult to adapt due to their continuous dynamics.
Furthermore, the efficiency gains of event-driven simulations can diminish in large, densely connected networks with high overall spike rates. In such regimes, the overhead of managing a vast number of concurrent events might negate the computational savings from only updating active elements. Nonetheless, as processing of synaptic events dominates computational requirements, the computational cost of emulating an entire brain can still decrease by a few orders of magnitude.
Figure 30 - Estimated Compute Requirements for Event-Driven Brain Emulation. The computational power (FP16 FLOPS) required for the same whole-brain emulations, but implemented using an event-driven paradigm (green bars), compared with reference hardware (grey). The y-axis is on a logarithmic scale. Error bars reflect the range of requirements based on the same varying assumptions about biophysical model complexity used for the other estimates. Calculations and underlying data for this figure are available in the [linked data repository](https://docs.google.com/spreadsheets/d/1so1BojMiuSTadc6BvGAkQL-CM9ieRaUkmFYWh4iGmYc/edit?gid=162927166#gid=162927166).
### Machine-readable figure record: `main-fig-30`
```yaml
id: main-fig-30
document_id: sobe-2025-main-ai
document_path: report.md
anchor: main-fig-30
canonical_url: https://brainemulation.mxschons.com/ai/report.md#main-fig-30
figure: "Event-driven emulation compute"
assets:
- "../assets/report/main-fig-30-publication.png"
data_status: final_circulation_export_raster_with_contextual_current_sources
generator: ../code/run_all_figures.py
generator_function: generate_emulation_requirements
publication_export_members:
- "images/image46.png"
canonical_pdf_pages:
- 160
```
#### Inline contextual data (not an exact publication input): hardcoded estimates
```tsv
organism mid_fp16_flops error_fp16_flops
C. elegans 5.27e8 5.23e8
Fly 2.61e11 2.50e11
Mouse 1.69e14 1.42e14
Human 6.04e16 2.64e16
reference:H100 1e15 0
reference:xAI_Colossus 1e20 0
```
#### Figure-generation source map (valid Python; not standalone plotting code)
```python
GENERATOR_PATH = "../code/run_all_figures.py"
SOURCE_FUNCTION = 'generate_emulation_requirements'
SELECTED_SOURCE_LINES = (
'event_mids = np.array([5.27e8, 2.61e11, 1.69e14, 6.04e16])',
'event_errs = np.array([5.23e8, 2.5e11, 1.42e14, 2.64e16])',
'ax.bar(x_pos_org, event_mids, yerr=event_errs, alpha=0.85, color=PURPLE,',
)
```
This can be seen in our estimates. Even in the more complex scenario, where the 5-compartment Hodgkin-Huxley neurons continue to be time-based, overall computational requirements still decrease by several orders of magnitude for all organisms considered, as a result of the more efficient event-driven modeling of synapses. It is essential, however, to contextualize these event-driven estimates within the current framework landscape. While GPUs and TPUs are now starting to be adopted by the computational neuroscience research community, enabling significant speedups over traditional CPU-based platforms like NEURON, and powering new frameworks such as ARBOR ([Akar et al., 2019](https://arxiv.org/abs/1901.07454)), BrainPy ([Wang et al., 2023](https://elifesciences.org/articles/86365)) and Jaxley ([Deistler et al., 2024](https://doi.org/10.1101/2024.08.21.608979)), their adoption is a relatively recent phenomenon, and remains limited. Further, event-driven primitives continue to be mostly lacking. Nevertheless, these trends suggest that, with sufficient research and development, even real-time emulation of the human brain may be feasible in the not-too-distant future, although likely requiring, at least initially, large-scale clusters.
------------------------------------------------------------------------
## Glossary
### A-C
Action Potential: A rapid rise and fall in voltage across a cellular membrane; the fundamental unit of communication for spiking neurons.
Barcoding (Protein/DNA): A technique (e.g., Brainbow, Tetbow, PRISM) used to assign unique molecular identifiers to individual neurons. This allows for the identification of disconnected neuron fragments in connectomics, potentially bypassing the need for perfect morphological tracing.
BOLD Signal (Blood Oxygen Level Dependent): The signal measured by fMRI, reflecting changes in blood flow and oxygenation that serve as a slow, indirect proxy for neural activity.
Calcium Imaging: An optical recording technique that uses fluorescent indicators (like GCaMP) to visualize the influx of calcium ions into a neuron, serving as a proxy for neuronal spiking activity. It allows for simultaneous recording of large populations but has lower temporal resolution than electrophysiology.
Connectome: The comprehensive map of neural connections (synapses) within a brain or nervous system.
Connectomics: The field of study focused on reconstructing the wiring diagram of the brain, typically using Electron Microscopy (EM) or Expansion Microscopy (ExM).
### D-F
Dense Reconstruction: The analysis and mapping of all neurons and their connections within a specific volume of tissue, as opposed to sparse reconstruction which maps only a subset.
Effectome: A quantitative map of causal influence between neurons, derived from perturbation experiments (like optogenetics) rather than just anatomical connections.
Electron Microscopy (EM): A microscopy technique using electron beams to achieve nanometer-scale resolution, essential for visualizing synapses. Variations include TEM (Transmission), SEM (Scanning), and FIB-SEM (Focused Ion Beam).
Electrophysiology: The branch of neuroscience that measures the electrical activity of neurons, ranging from single-cell Patch Clamps to large-scale Microelectrode Arrays (MEAs) like Neuropixels.
Embodiment: The integration of a brain model with a physical (robot) or virtual body that provides sensory input and executes motor commands, creating a closed sensorimotor loop.
Emulation: A computational model that matches a target system’s outputs by implementing the same internal causal dynamics at a chosen level of biophysical detail (distinct from Simulation).
Expansion Microscopy (ExM): A technique that physically expands biological tissue using swellable hydrogels to allow sub-diffraction limit features (like synapses) to be imaged using standard light microscopy.
fMRI (Functional Magnetic Resonance Imaging): A non-invasive imaging technique that measures brain activity by detecting changes associated with blood flow.
### G-L
Gap Junction: A specialized intercellular connection that allows various molecules, ions, and electrical impulses to pass directly between the cytoplasm of two cells (electrical synapse), often difficult to resolve in standard connectomics.
GCaMP: A family of genetically encoded calcium indicators (GECIs) derived from green fluorescent protein (GFP) and calmodulin, used to visualize neural activity.
GEVI (Genetically Encoded Voltage Indicator): Fluorescent proteins engineered to sense changes in membrane potential directly, offering faster temporal resolution than calcium imaging but currently suffering from lower signal-to-noise ratios.
Hodgkin-Huxley Model: A biophysically detailed mathematical model that describes how action potentials in neurons are initiated and propagated using explicit ion-channel kinetics.
In silico: Performed on a computer or via computer simulation.
Leaky Integrate-and-Fire (LIF): A simplified neuron model that treats the membrane as a resistor-capacitor circuit which spikes when a threshold is reached. It is computationally efficient but lacks detailed biophysical dynamics.
### M-P
Microelectrode Array (MEA): Devices containing multiple microscopic electrodes (e.g., Neuropixels) used to record neural signals from many neurons simultaneously.
Minimal Brain Emulation: A theoretical baseline defined in this report describing the minimum properties necessary for a model to be considered an emulation. Criteria include: \>95% neuron coverage, accurate synaptic connectome, cell type diversity, and consistent spiking activity scale.
Neuromodulators: Chemical messengers (e.g., dopamine, serotonin, neuropeptides) that regulate diverse populations of neurons over slower timescales than neurotransmitters, often diffusing through the extracellular space (volume transmission).
Neuropixels: A state-of-the-art high-density electrophysiology probe capable of recording hundreds to thousands of individual neurons simultaneously with millisecond precision.
Optogenetics: A biological technique that involves the use of light to control cells in living tissue, typically neurons, that have been genetically modified to express light-sensitive ion channels (opsins). Used for perturbation experiments.
Patch Clamp: An electrophysiological technique that allows the study of single or multiple ion channels in cells by sealing a glass electrode to the cell membrane. It is the gold standard for measuring single-cell electrical properties.
Perturbation: The deliberate manipulation of neural activity (e.g., via optogenetics) to establish causal relationships between neurons, as opposed to passive observation.
Photobleaching: The loss of fluorescence in a sample due to prolonged light exposure, limiting the duration of optical recording sessions.
Point Neuron: A simplified computational model where the spatial geometry of the neuron (dendrites, axon) is ignored, and the cell is treated as a single point in space.
Proofreading: The manual or semi-automated correction of errors (splits and merges) in the algorithmic neuron tracing of connectomics data.
### S-Z
Segmentation: The computational process of identifying and tracing the boundaries of neurons and organelles within raw microscopy image data.
Simulation: A model that matches a target system’s outputs (given the same inputs) without necessarily reproducing the internal causal dynamics or biophysical details (distinct from Emulation).
Synchrotron X-ray Tomography: A technique using high-energy X-rays to visualize large tissue volumes at high speeds. While currently limited in resolution compared to EM, it allows for non-destructive imaging of whole brains.
Transcriptomics: The study of the transcriptome (the complete set of RNA transcripts), often used to classify neurons into specific cell types based on gene expression.
Two-Photon Microscopy: An optical imaging technique that uses infrared light to excite fluorescent dyes. It penetrates deeper into tissue and causes less damage than single-photon microscopy, making it standard for in vivo recording.
Voltage Imaging: An optical method to measure the changing electrical potential of a cell membrane directly, capable of resolving individual spikes and sub-threshold events.
Volume Electron Microscopy (vEM): A set of EM techniques (like FIB-SEM or SBF-SEM) designed to capture 3D volumes of tissue rather than single 2D slices.
Whole Brain (Report Definition): A quantitative threshold defined in this report as a model incorporating at least 95% of neurons and at least 95% of the brain volume for a respective organism.