Chinese researchers have developed a new phase-change memristor chip that claims to run specific brain-mapping workloads up to 478.18 times faster than Nvidia’s A100 GPU, potentially opening new paths for brain-computer interfaces, digital brain twins, and neurological disease research.
TL;DR
- The chip was developed by Peking University and the Chinese Academy of Sciences’ Shanghai Institute of Microsystem and Information Technology.
- It uses controllable in-memory computing to reduce data movement and latency.
- The 478x claim applies only to specific brain-surface reconstruction workloads, not general AI training or inference.
Chinese researchers have revealed a new neurodynamic computing chip that claims a major leap in real-time brain modelling.
The team, led by Peking University professor Yang Yuchao, worked with researchers from the Chinese Academy of Sciences’ Shanghai Institute of Microsystem and Information Technology. Peking University said the work was published in Science under the title “A sub-10-millisecond neural dynamical system based on phase change memristors.”
The chip is based on phase-change memristors and uses a controllable in-memory computing approach. In simple terms, it reduces the need to constantly move data between separate memory and processing units, which is one of the biggest bottlenecks in conventional computing architectures.
That architectural shift is central to the performance claim. According to Peking University, the chip compresses a single-step neurodynamic system operation to 2.12 milliseconds and achieves 50.38 to 478.18 times faster performance than Nvidia’s A100 GPU in high-fidelity brain-surface reconstruction tasks. It also reported 3.82 to 36.27 times faster performance than advanced dedicated ASIC accelerators and 11.75 to 24.73 times lower power consumption.
The same claim was also reported by Chinese state outlet Xinhua, which said the chip uses a 40-nanometer process and integrates the two major compute tasks into a 0.28-square-millimeter in-memory computing array.
Topics For More Insights
Why Does This Matter?
Brain modelling workloads are different from ordinary AI workloads. Reconstructing the folded surface of the cerebral cortex requires repeated calculations, error control, and adaptive step-size searches. Traditional chips can do this, but they often lose time and energy moving intermediate data back and forth.
The new chip tackles that by using the physical behavior of phase-change memory itself. Peking University said the system maps integration steps and neural-network matrix operations directly into the device array, allowing the chip to perform parts of the calculation inside memory.
Yang said the breakthrough could support brain-computer interfaces and brain disease diagnosis. “In the future, personalised and dynamic digital brain twins will become possible,” he said, according to South China Morning Post.
The possible applications include brain-computer interfaces, real-time brain-state modelling, brain digital twins, intraoperative neuronavigation, and early screening for neurodegenerative diseases such as Alzheimer’s and Parkinson’s. Peking University said the chip may help future BCI systems move from basic signal recognition toward real-time brain-state modelling and intelligent interaction.
However, the result should not be read as a broad replacement for Nvidia GPUs. The comparison is based on the Nvidia A100, a data-center GPU introduced in 2020, and the 478x figure applies to a narrow brain-surface reconstruction workload rather than general AI model training, inference, graphics, or enterprise AI workloads. NDTV Profit also noted that the chip is designed specifically for neuroscience applications and is not meant to replace general-purpose GPUs.
Still, the work is significant because it shows how specialized chip architectures can outperform general-purpose GPUs when the workload is highly specific and data movement is the main bottleneck.
For now, the breakthrough is best seen as a research milestone in brain-inspired and in-memory computing. The next big questions will be whether the system can be independently reproduced, scaled, and supported with practical software tools for researchers and clinicians.
