Neuromorphic Processing Unit CereChip X1 Deep Dive: When Chips Begin to Mimic the Brain's Computing Style

Full Article

In May 2028, the Institute of Computing Technology at the Chinese Academy of Sciences (CAS) released the CereChip X1 neuromorphic processing unit in Beijing. This chip employs a compute-in-memory architecture that embeds computing units directly into storage arrays, fundamentally eliminating the bottleneck of frequent data shuttling between processor and memory in traditional von Neumann architectures.

Chen Yunjie, the CAS researcher leading the CereChip X1 project, explained at the launch event that the chip integrates 1.28 million analog neurons and 1 billion synaptic connections, natively supporting spiking neural networks (SNN). Unlike traditional GPUs running artificial neural networks, SNNs more closely approximate how biological brains work—neurons communicate through discrete spike signals rather than continuous value propagation.

In standard image recognition tests, CereChip X1 consumed only 0.3 millijoules per frame, one-eightieth of an NVIDIA H100 GPU. In time-series data processing tasks such as speech recognition and anomaly detection, its energy efficiency advantage was even more pronounced, reaching 120 times that of GPUs.

"The core value of neuromorphic computing lies not in speed but in energy efficiency," Chen explained. "For edge devices, IoT sensors, and mobile terminals, power consumption is often more critical than compute power. CereChip X1 can enable complex AI inference on button-battery-powered devices lasting years."

On the industrial application front, CereChip X1 has launched pilot projects with three partners. Hikvision is using it for real-time object detection in smart cameras, Huawei is integrating it into next-generation routers for network traffic anomaly analysis, and CATL is exploring its use in thermal runaway prediction for battery management systems.

However, the neuromorphic computing ecosystem remains immature. Current mainstream deep learning frameworks like PyTorch and TensorFlow are designed for traditional artificial neural networks with limited SNN support. CAS has open-sourced the companion SNN compiler SpikeFlow, but the developer community is still small.

Tsinghua University microelectronics professor Wei Shaojun noted that the biggest challenge facing neuromorphic chips is not the technology itself but the paradigm shift in programming. "Getting engineers accustomed to PyTorch to write spiking neural network programs involves a very steep learning curve. It requires restructuring the entire toolchain."

In the global competitive landscape, Intel's Loihi 2 and IBM's NorthPole remain leaders in neuromorphic chips. CereChip X1's release puts China in the first tier of this race, but industrialization will require continued investment in toolchains, developer ecosystems, and application scenario validation.

Chen Yunjie said the team is already developing the next-generation CereChip X2, targeting 10 billion synaptic connections at the same power consumption—"approaching the synaptic scale of a mouse brain."