news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Author | Lin Qingqing

Editor | Yuan Silai

At the end of 2022, the emergence of ChatGPT heralded the arrival of the era of big AI models.

The penetration of AI big models into various industries has stimulated the growth of the computing power market, and my country's intelligent computing power has maintained rapid growth. According to IDC, China's total intelligent computing power will reach 1117EFLOPS in 2027, with an average annual compound growth rate of 33.9%.

As the core component supporting computing power, AI computing chips are the computing power cornerstone of the cloud computing industry. The rapid construction of data centers, intelligent computing centers, and supercomputing centers in various parts of China in recent years has provided support for the research and development of the AIGC industry and multi-industry applications, and has further promoted the production and deployment of AI chips. According to the calculation of iResearch, by 2025, the market size of China's core artificial intelligence industry will reach 400 billion yuan, of which the market size of basic layer chips and related technologies will be about 174 billion yuan.

Hanxu Technology, which Hard Krypton recently contacted, is a computing chip research and development company based on the new paradigm of magnetic computing. It was incubated by the Center for Applied Magnetism of the School of Physics of Peking University. The company has recently completed two rounds of financing, seed round and seed + round. The investors are Zero Venture Capital and angel investor Fei Gesheng, and Yuanhe Capital serves as financial advisor.

Zhu Xinyue, CEO of "Hanxu Technology", told Hard Krypton that the team relies on Peking University's Department of Magnetism, which has decades of accumulation in magnetic principles, magnetic materials, and micro-nano magnetic devices. It is expected to iterate magnetic storage and develop the next generation of "magnetic computing" technology - "magnetic probability computing" and "magnetic logic computing" through micro-nano processing technology regulation, chip design, and process joint optimization.

At present, the AI ​​field does not lack computing power, but lacks available and easy-to-use computing power. Zhu Xinyue believes that "at present, all parts of the country are ushering in a new wave of enthusiasm for building intelligent computing centers. With the restrictions of US export controls and entity lists, the development of domestically controlled new computing chips is the key to solving the domestic AI computing bottleneck. The current high-bandwidth computing power requirements of large AI model reasoning and the mismatch with traditional chips are becoming more and more serious, calling for the emergence of new computing devices and computing paradigms."

In response to market demand, "Hanxu Technology" has delivered a new solution and launched the probabilistic computing chip SpinPU-M01, a new AI computing chip based on spin. The computing power it provides can be quickly deployed in drug development, scientific computing, air logistics and other fields for dedicated computing acceleration; in addition, it has developed the SpinPU-E series of magnetic logic computing chips, which use a lower-cost on-chip MRAM high-bandwidth solution to provide inference computing chips for edge and cloud large models.

Probability computing chip SpinPU-M01

In terms of chip material system, Hanxu Technology uses CMOS-compatible magnetic chips instead of mainstream traditional silicon-based chips. Although the process technology of the former is relatively mature, the advanced process is blocked in China and difficult to obtain. The latter can achieve large-scale on-chip storage due to the superior characteristics of magnetic devices, and can provide the ultra-high bandwidth required for AI reasoning; secondly, magnetic chips have a higher degree of integration and can realize larger-scale storage units in a limited area.

Compared with quantum computing, probabilistic computers do not require harsh conditions such as ultra-low temperatures close to absolute zero (-273.15℃). They can work at room temperature and are superior to adiabatic quantum computing in terms of cost and larger-scale computing, making them closer to practical use.

Compared with traditional computing power, in solving combinatorial optimization problems, different problem scales, that is, different computational complexities, will have the same solution time scale when deployed on adiabatic quantum computers and probabilistic computers. Magnetic computing chips can achieve the same performance at the 28nm process node with device advantages and memory access speed advantages, at a lower cost, and large-scale integration can be achieved with domestic processes.

Zhu Xinyue told Hard Krypton that the latest product SpinPU-M01 has verified the obvious advantages of probabilistic computing in accelerating AI pharmaceutical manufacturing in usability testing, stability testing, and authoritative data set testing in industrial-grade environment testing of protein docking. Compared with the main series' highest-performance CPU Intel i9-14900k chip released in Q3 2023, the performance is at least 85.81 times higher.

Regarding core advantages, Zhu Xinyue believes that "in the research and development of annealing machines with larger bit numbers, the 'spin solution' has lower costs, does not require ultra-low temperatures, and has higher integration than the 'superconducting solution' and the 'photon solution', and can quickly expand annealing machines with larger bit numbers. And compared with the large-scale quantum complete equipment solution, our chips and accelerator cards will use a universal PCIe interface, which makes it easier to expand the cloud computing power of intelligent computing centers in various places and the edge computing power of terminal-side scenarios."

Hard Krypton learned that in terms of product progress, "Hanxu Technology" has released the 1024 probability bit fully connected SpinPU-M01 chip and acceleration card.

Concept diagram of the logic computing chip SpinPU-E01

It is expected that in 2025, the SpinPU-E series of magnetic logic computing chips will be launched, which will be customized for the Transformer architecture and have the advantages of low power consumption, high bandwidth and high speed. It is expected to adapt to mainstream AI models and complete terminal deployment within limited power consumption.

It is reported that the storage bandwidth density and inference energy efficiency of the SpinPU-E series chips will be directly comparable to NVIDIA H100 (TSMC 4nm), and the peak computing power density will be close to NVIDIA Orin (TSMC 7nm).