China Launches First Commercial 28nm Memory-Compute AI Chip

TMTPOST--Beijing-based chip design company PIMCHIP on Thursday announced the launch of its PIMCHIP-N300 and PIMCHIP-S300 AI chips, marking a milestone as China's first commercially available 28nm memory-compute AI chips.

These chips integrate memory and computation capabilities and support a range of applications including smart wearables, intelligent security systems, embodied intelligence, AI large models, and health data analysis.

The PIMCHIP-N300 is designed as a next-generation neural network processing unit (NPU) optimized for machine learning and AI applications. It supports mixed-precision computing, including 4-bit, 8-bit integer, and 16-bit floating-point operations.

This chip aims to balance power consumption, computational density, and precision, providing a complete development toolkit including NPU intermediate representation layer specifications, model parsers, optimizers, and drivers. The N300 is expected to accelerate product development and meet diverse customer needs.

The PIMCHIP-S300 is a high-efficiency, compact, low-power AI chip developed for real-time control and scheduling. It integrates a lightweight MCU processor and supports multi-modal sensing with audio and video inputs.

The S300 features advanced digital PIM units and a proprietary heterogeneous architecture, capable of ultra-low-power wake-up, voice activity detection, speech recognition, motion monitoring, and visual recognition. This chip can achieve up to 90% energy savings for specific computational tasks, with a core energy efficiency of 27 TOPS/W, making it ideal for smart wearables, embodied intelligence, and AI large models.

Yang Yue, the co-founder and CEO of PIMCHIP, highlighted that the rapid advancement of chips and AI technology is reshaping various industries. He emphasized that the company’s new products would drive transformations from smart devices to smart cities and from intelligent healthcare to embodied intelligence.

Yang expressed optimism about collaborating with leading industry players to explore the practical applications and value of memory-compute integration technology.

Yang also shared that PIMCHIP has already begun collaborating with leading domestic and international electronics companies, providing one-stop solutions for in-memory computing technology. The company is progressing towards 22nm and 14nm technology nodes, with growing customer demand for these advanced solutions.

AI 1.0 is maturing, particularly in neural network computations supported by the newly launched products. AI 2.0, with its broader applicability, is expected to transform various fields such as image and video generation, he added. Yang believes that while commercialization might take time, technological progress will continue, similar to how electric vehicles have evolved over the years.

Founded in February 2021, PIMCHIP positions itself as an innovator in intelligent computing architecture, dedicated to providing technological support for the widespread application of AI through innovative memory-compute solutions. The company has applied for over 40 patents domestically and internationally.

Wang Yu, a professor at Tsinghua University and a long-term collaborator and friend of Yang, praised the new chips for their performance improvements, versatility, and platform architecture.

As the memory-compute technology enters the commercialization phase, major tech giants like Nvidia, Microsoft, and Samsung are pumping money into this field. In China, several startups and listed companies, including ZBIT SEMI Co., Houmo AI, and Witin Tech, are focusing on this technology, with products moving towards mass production.

China’s market for memory-compute technology is estimated to hit 12.5 billion yuan by 2025, with expectations of growing to 113.6 billion yuan by 2030 as technology matures and large-scale commercial applications become more widely adopted.