Peking University researchers turned a phase change memory flaw into a computational feature, beating a 2020 vintage Nvidia A100 on cortex reconstruction, though only on that one task.
Researchers at Peking University and the Chinese Academy of Sciences [published a peer-reviewed paper in Science on July 4, 2026](https://www.scmp.com/news/china/science/article/3359408/chinese-scientists-brain-mimicking-chip-478-times-faster-nvidia-a100-gpu) showing a chip that can reconstruct a rat brain's folded cortex in under half a second, 478 times faster than a 2020-vintage Nvidia A100. The mechanism behind that number, a phase-change memristor array that turns a memory defect into a computational feature, is the actual contribution. The work was picked up by the South China Morning Post, the Seoul Economic Daily, and Chinese state broadcaster CGTN.
The architecture is built on phase-change memristors, memory elements that store data in their resistance and switch states with heat. A phase-change memristor normally has a stubborn flaw: its conductance drifts after programming, which makes precise analog computation unreliable. The Peking team, led by professor Yang Yuchao, modeled that drift mathematically and used the model as a controllable parameter, then built a 0.28 square-millimeter array that does matrix multiplication and accumulation inside the memory itself.
Conventional computers shuttle data between a separate processor and memory, the von Neumann bottleneck that has shaped chip design since 1945. Compute-in-memory, which runs the math where the data is stored, collapses the trip. On the team's specific benchmark, reconstructing the folded surface of a rat brain cortex, the chip hit 2.12 milliseconds per iteration and beat a 2020-era Nvidia A100 by 50x to 478x, depending on configuration. On related neural-dynamics tasks it was 3.82x to 36.27x faster than dedicated accelerators.
Two researchers at Germany's Jülich Research Center, writing in the same issue of Science, called the approach "processing crude oil directly at the oil field rather than transporting it to a factory." That is the architectural argument: the cost in time and energy to move data to a separate processor is, for this workload, a much bigger penalty than the raw math would suggest.
The chip is still a task-specific accelerator, not a general-purpose GPU replacement. The A100 launched in 2020; Nvidia has since shipped the Hopper H100 and the Blackwell B200, both of which are fenced off from China by US export controls introduced in October 2022. The Peking team compared its chip against a part two generations out of date and on a single task: folded-cortex surface reconstruction. Yang, quoted in Chinese state media, said the chip was designed for brain-computer interfaces, Alzheimer's screening, intraoperative neuronavigation, and personalized digital brain twins. All are workloads where latency and energy per inference matter more than peak throughput.
The defect-as-feature idea will be tested on larger arrays, smaller process nodes, and workloads outside brain modeling in the next eighteen months. Conductance drift has long been treated as a defect to be engineered around in phase-change memory. The Peking work turned the drift into the central computational primitive. The 40nm node and the 0.28 square-millimeter array are small. If the trick survives contact with scale, analog compute-in-memory with calibrated, drifting devices starts to look like a genuine alternative path for low-latency inference at the edge.