Liangxun Technology has completed its angel round financing to build an AI acceleration and quantum-inspired computing platform.

Recently, Xunwei Technology announced the completion of its angel round of financing, which was led by Peak Ventures, with joint participation from Pudong Venture Capital and Chengding Fund under Shanghai Chengtou Group. The funds raised in this round will be mainly used for the tape-out of core products, technical R&D, team building and market expansion, to accelerate the R&D and commercialization of PPU (Probabilistic Processing Unit) and AI inference acceleration chips. Duowei Capital acted as the exclusive financial advisor for this new round and for the long term.

Xunwei Technology is positioned as AI acceleration and quantum-inspired computing platform, and develops along two product lines: one targets the fast-growing demand for AI inference, and the other promotes the processorization of probability, search and optimization computing through PPU (Probabilistic Processing Unit). The two lines share the underlying device and chip engineering capabilities accumulated by Xunwei over a long period of time, and also correspond to the company's judgment on the next stage of the computing industry: The demand for computing power from AI is gradually extending from single-model computing to more types of workloads, and quantum-inspired computing is also looking for a window to enter the real industry earlier.

The problems faced by AI have changed, and the requirements for underlying computing are also evolving

Over the past decade, one of the most important tasks of AI computing power is to make matrix computing faster and faster. From GPUs to various AI accelerators, this system has supported the explosion of deep learning and large models all the way. However, as AI further evolves from content generation to Agent, embodied intelligence and real-time decision-making, computing tasks have become more diverse.

An Agent needs to choose between different tools and actions, a robot needs to plan paths in real time according to the environment, and more complex AI systems need to continuously search and optimize among a large number of possibilities. AI is increasingly facing the problem of "how to choose the next step", corresponding to computing workloads such as search, sampling, optimization and probabilistic decision-making.

For Xunwei, AI inference acceleration will solve the problem of efficient model execution, while probabilistic computing further addresses the new computing demands arising from the extension of AI to search, planning and decision-making.

This change is also spawning more computing chips redesigned for specific workloads. In March, Samsung Catalyst Fund led a $50 million financing round for Normal Computing, which explores computing using the inherent randomness of physical systems. In July, Extropic signed a letter of intent for up to $75 million in CHIPS R&D funding with the U.S. Department of Commerce to promote the industrial implementation of probabilistic computing units. In August, AMD announced the acquisition of Taalas, a dedicated AI inference chip company, which tries to deeply solidify the data flow and computing structure of models into dedicated silicon chips.

In China, probabilistic computing coprocessors have also been included in the scope of attention of local future industries and new computing-related projects, and Xunwei has recently received support from government projects.

As AI workloads continue to diversify, underlying hardware is also extending from general-purpose computing power to more segmented computing tasks. In addition to efficient inference, whether workloads such as search, sampling, optimization and decision-making can form more suitable hardware implementation methods has also become part of the new round of computing power innovation.

Quantum-inspired computing, looking for a closer implementation window

On the other hand, the development of quantum computing and statistical physics also provides a technical idea for these search, optimization and complex solving problems that is different from traditional computing. One direction is to transfer the ideas of dealing with complex problems in quantum computing and statistical physics to the software and hardware systems that can be manufactured and deployed today, to take the lead in solving problems with high traditional computing costs such as search and optimization.

Xunwei chose PPU, hoping to use the classic semiconductor manufacturing system to transform the computationally valuable capabilities in quantum and statistical physics into engineerable chips.

Since the beginning of this year, relevant industrial actions around the world have increased significantly. In February, Toshiba and MIRISE Technologies applied quantum-inspired optimization to real-time decision-making of autonomous mobile robots. In June, Toshiba further combined AI with quantum-inspired optimization, targeting dynamic optimization scenarios such as automobiles, wireless communications and robots. In September, D-Wave signed a funding support agreement of up to $100 million with the U.S. Department of Commerce to promote the R&D and large-scale application of quantum-inspired computing.

Different technical routes point to the same question: Can the randomness in the physical world be further transformed from "noise" that needs to be suppressed in traditional computing into available computing resources?

AI has brought more and more computing demands for search, sampling, optimization and decision-making, and quantum-inspired computing is looking for an earlier industrial implementation window. Xunwei aims at the intersection of the two technical contexts.

One underlying technology, two computing routes

Seeing the trend is one thing, finding a technical path that can really make chips is another.

Xunwei's answer comes from the underlying device technology accumulated by the team for many years. The same spin device can work in two states: deterministic and probabilistic, thus corresponding to two different types of computing demands.

In the stable state, the device can be used for storage and computing, serving AI inference through in-memory computing; after entering the probabilistic state, it can use the inherent randomness of the device to participate in computing for tasks such as sampling, search and combinatorial optimization.

One set of underlying devices thus extends to two product lines. AI inference acceleration chips target end-side scenarios such as machine vision, intelligent driving and embodied intelligence, to undertake the rapidly growing demand for computing power. PPU targets probabilistic computing and tasks such as sampling, search and combinatorial optimization, to precipitate new hardware computing capabilities.

Why Xunwei?

For new computing technologies, the longest distance in the industrialization process is from spin devices, circuit design to chip tape-out, algorithm mapping and final application. The Xunwei team has accumulated along this chain for many years, and has promoted the underlying device capabilities to the stage of chip engineering and application verification.

Dr. Zhu Zheng, CEO of Xunwei Technology, has more than 10 years of industrial experience in chips, quantum and security, and is an expert in spin device technology. Professor Wang Jiayong, Chairman, has nearly 30 years of experience in the quantum and chip industry and industrialization. Hou Yaoru, CTO, has long focused on spin memory and probabilistic computing, and has experience in mass production verification of tens of thousands of related chips. Academician Xu Hongxing, Chief Scientist, is an academician of the Chinese Academy of Sciences, with long-term research accumulation in the field of quantum and chips. The core R&D team gathers technical backbones from universities such as the Hong Kong University of Science and Technology, Beihang University, Shanghai Jiao Tong University, Fudan University and ShanghaiTech University, forming a complete R&D echelon from devices to systems.

The scarcity of this team lies in its simultaneous capabilities in device, chip and industrialization.

Previously, the Xunwei team has completed the full verification of spin devices from process, tape-out, packaging and testing to industrial application in quantum security related products, and accumulated a set of reusable device engineering and chip mass production capabilities. This enables the company to promote AI inference acceleration chips and PPU today not to start from scratch in the laboratory, but to continue to extend to new computing scenarios on the basis of existing engineering.

At present, Xunwei's third-generation AI inference acceleration chip and PPU have entered the tape-out stage, and are carrying out return chip testing, continuing to iterate in chip architecture, algorithm mapping and application systems.

In terms of commercialization, the company's AI inference acceleration chip has carried out full-stack "hardware + software" cooperation with the intelligent driving department of a well-known domestic automaker. PPU is promoting the transformation of algorithm scheduling to computing systems, and discussing the technical path of heterogeneous coprocessors with leading domestic high-performance GPU manufacturers.

As AI workloads continue to diversify, there may be more and more dedicated computing units designed for different tasks in the future computing system, each playing a role in the most suitable computing problems.

When AI begins to learn how to make choices, computing also needs to learn to deal with uncertainty.

What Xunwei wants to make is that earlier emerging chip.

Peak Ventures said: "Quantum technology is a key national strategic direction and an important part of the new round of scientific and technological revolution. Based on the silicon-based electronic spin platform technology, Xunwei has taken the lead in realizing the tape-out, mass production and customer application of quantum random number generator chips, and further expanded application fields such as AI inference acceleration, quantum security and quantum-inspired computing. We are optimistic about the team's capabilities in chip design and industrialization, and expect Xunwei to further deepen the silicon-based electronic spin technology into larger-scale chips and real computing system applications."

Pudong Venture Capital said: "The artificial intelligence industry is still in the early stage of rapid evolution of technology and applications, and computing power is particularly important as the underlying basic capability. Xunwei's simultaneous layout of AI inference and probabilistic computing shows us the possibility of extending new computing power to more application scenarios. We believe that as AI enters the real physical scene, the demand for underlying computing capabilities will continue to expand, and Xunwei has the opportunity to grow into a hard technology company with long-term industrial value in this process."

Chengding Fund said: "In addition to the technology itself, we also pay close attention to how new technologies are implemented. Chips are ultimately verified by products and customers. Xunwei has now entered the stage of return chip testing, and has begun to work with industrial customers on applications. What follows is a very practical matter: to make good products and get through the scenarios. Shanghai has rich industrial and application resources, and we will try our best to help the company connect these resources, so that the technology can reach the real market faster. "