SpiNNaker and what a million real time spiking cores can and cannot test
SpiNNaker is a supercomputer built to run spiking neural networks in real time, and it is the clearest hardware test of a specific consciousness hypothesis. Its full machine, assembled at the University of Manchester, holds 57,600 nodes and 1,036,800 ARM cores with over 7 TB of memory, and it is designed to simulate on the order of a billion neurons in real time. The design goal is not raw speed but temporal fidelity, the machine must complete a neural computation fast enough that the simulated brain keeps pace with the biological clock.
The consciousness connection is the reason the machine exists. If consciousness is a property of dynamic organization, of recurrent spiking activity with the right temporal structure, then a machine that can run that structure at biological speed is the hardware on which the claim can be tested. SpiNNaker is that machine, and the question of what it can and cannot establish is the subject of this post.
What SpiNNaker Actually Is
SpiNNaker, short for Spiking Neural Network Architecture, is the work of Steve Furber’s Advanced Processor Technologies group at Manchester. Each of its ARM968 cores emulates up to 1,000 neurons using a bespoke communication fabric that moves spike events between cores rather than sharing a central clock. The architecture is event driven. When a neuron fires, the machine routes a small packet to the cores that need it, which is the same economy a biological brain uses and entirely unlike the dense matrix work a GPU does.
The machine reached its million-core milestone on October 14, 2018, and the achievement was announced by the Human Brain Project. Its successor, SpiNNcloud, built by TU Dresden under Christian Mayr, began operation in 2025. The successor keeps the spiking principle and adds the process technology a modern compute node needs.
The distinction from the site’s other neuromorphic coverage matters. Intel Loihi 2 is a chip you can program to run different neuron models. SpiNNaker is a machine-scale simulation platform, and its defining constraint is real time, computing fast enough that a simulated millisecond of brain time completes in a millisecond of wall time. That constraint is the whole point, because the theories the machine is built to test are theories about timing.
The Consciousness Connection
The question this post asks first is how SpiNNaker bears on consciousness, and the answer is a specific architectural claim. Consciousness theories built on temporal dynamics, recurrent processing theory as covered on this site and the global workspace broadcast account, predict that the pattern of neural activity in time, not its content alone, is what makes a computation conscious. Those theories cannot be tested on a GPU running a simulation slower than real time, because a slow simulation does not have the same temporal structure as the brain it models.
SpiNNaker closes that gap. A recurrent network of model neurons running at biological speed on SpiNNaker has, in principle, the timing that the theories stipulate. If a researcher wants to ask whether a particular recurrent architecture produces a workspace-like ignition or a recurrent-processing signature, SpiNNaker is the platform where the temporal condition is actually met.
What it does not establish is the harder half of the question. Running the dynamics at the right speed does not prove that the dynamics generate experience, and the substrate debate contests exactly this, whether biological ion-channel biophysics is constitutive or incidental. SpiNNaker is the strongest test of the functional-dynamics hypothesis available, and it is silent on whether satisfying that hypothesis is sufficient.
The Neuron Model Question
The site’s Izhikevich analysis documents why the choice of spiking model matters, a leaky integrate-and-fire neuron covers 3 of 20 cortical firing behaviors, while a two-variable model can be tuned to all 20. SpiNNaker’s cores can implement either, because the architecture is programmable, which means the same hardware can run the minimal LIF substrate or the richer Izhikevich dynamics. That programmability is what makes it a test platform rather than a fixed demonstration.
The comparison to the Substrate Console is direct. The console runs the site’s own LIF-based basal ganglia model in the browser. SpiNNaker is the machine that could run the same class of model at brain scale in real time, which is the difference between inspecting a small circuit and testing whether a full-scale version produces the dynamics the theories require.
The Real Time Constraint
Real time is a stronger constraint than it sounds. Modern simulation often trades temporal fidelity for throughput, running a small network much faster than real time or a large one much slower, and neither regime has the timing structure a temporal-dynamics theory cares about. SpiNNaker is built to hold that structure, and the cost is visible in its power draw, roughly 100 kilowatts for the full machine, which is high by compute standards and low relative to the alternative, a cluster that cannot hold real time at all.
That power cost connects to the orbit section’s argument. The power budget analysis and space solar post on this site treat sustained watts as a precondition for any long-horizon consciousness computation. SpiNNaker is a 100-kilowatt example of the compute term of that equation, and the fact that a billion-neuron real-time simulation needs that much power in silicon is a concrete data point for what the brain emulation section tracks.
What It Can and Cannot Test
What SpiNNaker can test is specific and valuable. It can test whether a given recurrent spiking architecture, at scale, at biological speed, produces the dynamic signatures that temporal-dynamics theories specify, workspace ignition, recurrent processing, gamma-band structure, cross-region phase locking. That is a real experiment and this hardware is the first place it can be run.
What it cannot test is whether those signatures, once produced in silicon, are accompanied by experience, because no machine measurement can decide that. The indicator approach treats the signatures as evidence, not proof, and SpiNNaker is the strongest generator of that evidence class that exists. The machine sits exactly at the boundary the site’s framework draws, between the substrate-independence claim, which it can test, and the phenomenal claim, which no hardware alone can.
Where This Sits
SpiNNaker is the emulation section’s temporal-fidelity anchor. It joins the whole brain emulation roadmap, which defines how much of a brain a copy would need, and the Moravec transfer, which asks what gradual replacement assumes. Those posts ask whether emulation is possible in principle and what it would mean. SpiNNaker asks whether the temporal dynamics the theories need can be run at all, which is the engineering precondition the other questions depend on.
The current scientific consensus remains the right frame for the whole question. SpiNNaker does not move the consensus by itself. It removes an excuse, showing that the dynamic side of the leading theories can be brought to real time in silicon, and that the obstacle to testing them is no longer hardware but the theoretical question of what the correct model and the correct signatures are.
SpiNNaker’s full machine holds 1,036,800 ARM968 cores across 57,600 nodes with over 7 TB of memory. The million-core milestone was announced October 14, 2018 by the Human Brain Project. SpiNNcloud, its second-generation successor at TU Dresden under Christian Mayr, began operation in 2025. Source: the University of Manchester SpiNNaker project and the Human Brain Project.
