according to media reports, shanghai zhongzixing technology co., ltd. has developed an intelligent natively-designed chip called neu (neural execution unit) and plans to begin tape-out in the fourth quarter of this year.
based on measured data disclosed by zhongzixing, under the same ai inference tasks, the neu chip achieves nearly 100 times the speed of leading conventional gpu solutions, while consuming only one-tenth of the power. unlike following nvidia’s “gpu-like” ecosystem chip design approach, zhongzixing has taken a different path: the neu is designed with memory at its core, eliminating the need for frequent data movement between storage and compute units, thereby fundamentally removing the energy overhead and bandwidth bottlenecks associated with data movement in the traditional von neumann architecture.
at the same time, in terms of hardware topology, the chip directly implements neuron-style connections through internal interconnects, natively supporting current mainstream neural network architectures. it has also been optimized for future fine-grained unstructured sparse models and irregular neuron topologies, achieving structural isomorphism with biological intelligence. zhai sitong, founder of zhongzixing, previously led the development of hisilicon’s entire eda software framework. in an interview, he stated that the core team possesses deep technical expertise and strong engineering capabilities, and that after founding the company, they have deeply integrated years of laboratory research results with large-scale industrial integration experience.
currently, zhongzixing has brought together interdisciplinary talent from intel, nvidia, hisilicon, and top canadian ai laboratories, enabling end-to-end in-house r&d spanning architectural design, compiler optimization, and physical-model training.
