The Japanese supercomputer Fugaku reproduces the activity of 10 million cortical neurons in 32 seconds to simulate a single second of actual thought. This performance, achieved by the Allen Institute and the University of Electro-Communications of Japan, marks the first complete simulation of the mouse cortex in quasi-real time. Neuromorphic chips promise to divide the energy consumption of these calculations by 1000.
Fugaku reproduces mouse brain activity in quasi-real time
The Fugaku supercomputer, installed at the Riken research center in Kobe, mobilizes 152,000 ARM processors to simulate the equivalent of a complete mouse cerebral cortex. The 10 million artificial neurons exchange 37 billion signals per second, faithfully reproducing the activity patterns measured on the living brain.
This simulation achieves a 32:1 ratio – it takes 32 seconds of computation to model one second of actual neural activity. A spectacular leap compared to previous attempts that required several hours for the same result. The Japanese team optimized synaptic transmission algorithms and parallelized calculations across Fugaku’s massively distributed architecture.
Precision reaches the level of the individual synapse. Each virtual neuron maintains its own electro-chemical characteristics: membrane potential, activation thresholds, transmission delays. The model even reproduces the gamma oscillations observed in the real cortex, these 40 hertz patterns that synchronize neural activity during cognitive processes.
The energy chasm between classical computing and biological brain
Fugaku consumes 30 to 40 megawatts at maximum operation to simulate a mouse brain that consumes a fraction of a watt (the human brain consumes 20 watts). This difference of approximately 1000-2000x reveals the inefficiency of digital processors compared to biological circuits optimized by 500 million years of evolution.
The brain processes information in an analog and massively parallel manner. Each neuron calculates simultaneously with its thousands of neighbors, with no distinction between memory and processor. Supercomputers, conversely, decompose each operation into sequential binary calculations, multiplying data transfers between computing units and memory.
This energy consumption drastically limits practical applications. The human brain contains 86 billion neurons (16 billion cortical), representing a considerable energetic and technical challenge for any complete simulation.
2.5 petabytes represents the estimated storage capacity of the brain, not its processing throughput. No current digital system approaches biological efficiency, even exascale supercomputers that reach the symbolic threshold of 10^18 operations per second.
Neuromorphic chips promise 1000 times greater efficiency
IBM, Intel, and Qualcomm are developing neuromorphic processors that directly imitate brain architecture. These chips abandon binary logic for continuous analog calculations, bringing silicon closer to synaptic functioning.
Intel’s Loihi 2 chip integrates 1 million artificial neurons on 31 square millimeters. Each neuron simulates the accumulation of electrical signals until the triggering threshold, exactly like its biological counterpart. Consumption drops to 1 microwatt per active neuron, 1000 times less than classical processors for the same calculations.
IBM’s TrueNorth pushes this logic further with 4096 neuromorphic cores connected by an asynchronous communication network. No global clock: each event instantaneously triggers the following calculations. This event-based approach reproduces the reactivity of the biological brain where only stimulated neurons consume energy.
BrainChip already commercializes the Akida, a production neuromorphic chip that processes image recognition in real time with 10,000 times less energy than a graphics processor. Learning occurs directly in the chip, without transfer to external memory. This innovation could revolutionize embedded artificial intelligence in autonomous vehicles and mobile devices.
The challenge of the complete human brain
The human brain contains approximately 16 billion cortical neurons, more than 1600 times the current simulation on Fugaku. This scaling up encounters fundamental obstacles that far exceed current computing capacity.
Synaptic connections grow exponentially with the number of neurons. Each human cortical neuron connects to approximately 7000 others on average, creating 150 trillion synapses in total. Simulating these connections exceeds the memory capacity of all existing computer systems.
Temporal complexity poses an additional challenge. The human cortex combines oscillations of multiple frequencies: delta (1-4 Hz) for deep sleep, alpha (8-13 Hz) for relaxation, gamma (30-100 Hz) for focused attention. These rhythms interweave across different time scales, from milliseconds to circadian cycles.
The University of Manchester is developing SpiNNaker 2, a neuromorphic supercomputer of 10 million ARM processors specialized in neural simulation. The objective: to achieve 1 billion simulated neurons by 2027, or 5% of a human brain. But reproducing synaptic plasticity – the adaptive capacity underlying learning – remains largely unexplored.
Medical applications open an immediate path
The precise simulation of the mouse cortex is already accelerating neurological research. Pharmaceutical laboratories virtually test the effect of molecules on neural circuits before preclinical trials, reducing development costs and timelines.
Epilepsy directly benefits from these models. Neurologists reproduce convulsive seizures in the virtual cortex to identify epileptogenic zones and test different therapeutic strategies. This approach could personalize treatments by simulating the effect of antiepileptic drugs on each patient’s specific brain.
The promise of personalized gene therapies against rare diseases could converge with these neural simulations. Modeling the impact of genetic mutations on brain activity would guide the development of targeted treatments for neurodegenerative pathologies.
Brain-machine interfaces are progressing thanks to these detailed models. Understanding precisely how neurons encode motor intentions improves neural prostheses that allow paralyzed patients to control robotic limbs through thought. Neuralink and Synchron rely on this knowledge to optimize their brain implants.
Neuromorphic supercomputers could transform real-time medical diagnosis. Analyzing EEG signals with the precision of a simulated brain would detect early signs of dementia or stroke, opening crucial therapeutic windows.
Sources: