Neural dynamical systems combine neural networks with equations describing how complex systems evolve, and they are computationally punishing: they demand repeated calculation, error checking and step-size adjustment. Conventional architectures compound the cost by moving data back and forth between memory and processor. A team led from Peking University with the Shanghai Institute of Microsystem and Information Technology built a chip that performs the key operations where the data already sits.
The device is built on a 40 nanometre process, with in-memory computing and conductance-drift arrays occupying 0.28 square millimetres. It runs at 50 megahertz and uses nine pipeline stages for each integration step. On neural dynamics calculations it proved between 3.82 and 36.27 times faster than state-of-the-art application-specific integrated circuits while consuming between 11.75 and 24.73 times less power, a rare combination of both advantages at once.
The demonstration chosen was cortical surface reconstruction, where the chip achieved up to a 478-fold speedup over a high-end graphics processor. The researchers used it to reconstruct the surfaces of the brain's white and grey matter and generate three-dimensional surface meshes in real time, producing output that was smooth, closed and topologically consistent while capturing complex cortical folding. The work was published in Science.
Real-time high-quality brain modelling has applications that batch processing cannot serve: brain-computer interfaces, digital brain twins, surgical navigation and medical imaging all depend on results arriving while the clinical situation is still live. The broader significance is as a demonstration that in-memory analogue computing can handle a genuinely demanding numerical workload rather than only the matrix multiplications of inference.