Overview
- Researchers from Peking University and the Chinese Academy of Sciences published a paper this week in Science describing a 40‑nanometre computing‑in‑memory chip that they say reconstructs the brain’s folded cortical surface in under 0.5 seconds.
- The team reports workload‑dependent speedups of roughly 50× to 478× versus an NVIDIA A100 GPU for the same cortical‑reconstruction tasks, but those numbers apply to the reported benchmarks and may vary with different workloads or implementations.
- The chip uses phase‑change memristors to store weights and perform calculations in the same hardware, and it deliberately exploits conductance drift — a device physics effect usually seen as a flaw — to carry out neural‑dynamical computation faster and with lower energy use.
- Independent experts from Germany’s Juelich Research Centre, in an accompanying analysis, highlighted millisecond‑scale latency that could support real‑time clinical uses, but they and the authors note the need for independent benchmarking, systems integration, regulatory review, and clinical validation before hospital deployment.
- This work builds on a broader push to solve the memory‑to‑compute bottleneck in high‑resolution brain imaging and, if validated and integrated into imaging and surgical workflows, could shorten the time patients wait for scans and enable faster intraoperative mapping and point‑of‑care screening.