The Rise of RISC-V as an AI-Native Computing Architecture
The recent surge in popularity of DeepSeek has sent shockwaves through the AI industry, and its influence extends far beyond the realm of artificial intelligence itself. The semiconductor industry, in particular, has taken keen notice. During the Chinese New Year, Alibaba’s DAMO Academy Xuantie announced its adaptation of the DeepSeek-R1 series distillation model, a powerful demonstration of the burgeoning momentum of the open-source instruction set architecture, RISC-V, within the AI domain.
At the recent Xuantie RISC-V Ecosystem Conference, exciting developments were unveiled: RISC-V has achieved significant breakthroughs in both high-performance computing and AI. The Xuantie C930, DAMO Academy’s first server-grade CPU, is slated to begin delivery next month. Its significantly enhanced AI computing power is poised to accelerate the deployment of a comprehensive ‘high-performance + AI’ RISC-V ecosystem.
This raises a crucial question: Could the open-source computing architecture RISC-V be the ideal partner for open-source AI?
AI Model Transformation Fuels Innovation in Computing Architecture
A seasoned expert in the chip industry explained that DeepSeek’s impact is felt not only within AI circles but also profoundly within the chip industry itself. DeepSeek, through its highly optimized design, has drastically reduced the training and inference costs associated with large language models. This shift has dramatically altered the existing balance of computing power, memory, and interconnection, creating significant opportunities for breakthroughs in computing architecture.
Traditionally, large AI models, owing to their intensive computing and memory requirements, were better suited for deployment in the cloud rather than on edge devices. However, DeepSeek’s arrival has challenged this reliance on high computing power. By substantially reducing both training and inference costs, it is paving the way for large models to transition from the cloud to the edge.
Specifically, DeepSeek’s reduced computational demands make single-machine deployment feasible, enhancing its compatibility with edge and end-side devices. As AI seeks to penetrate diverse industries and scenarios, the need to move from the cloud to the edge becomes increasingly critical. This shift is necessary to meet diverse needs such as data security, personalized customization, and private deployment.
It is foreseeable that, with the widespread adoption of DeepSeek technology, the landscape of AI chips will undergo a transformation. From large-scale parallel computing reliant on cloud infrastructure, AI chips are evolving towards diversified, efficient, and low-power designs capable of independent operation on edge devices.
This evolution has prompted many in the industry to ponder: what computing architecture is best suited for the demands of AI?
GPUs, with their inherent parallel processing capabilities, may not be the sole solution. Serial computing (general-purpose computing) is also emerging as a viable foundation for AI computation. Industry experience demonstrates that DeepSeek exhibits good compatibility with various computing systems. Its ability to be rapidly deployed and perform effective inference on CPUs has brought CPUs back into the spotlight. Compared to specialized GPUs, CPUs offer the advantage of versatility, simplified scheduling, a significant reduction in computing power requirements, and the benefits of homogeneous computing.
Among CPUs, the rising star, RISC-V, is attracting significant attention and investment.
During the Chinese New Year, DAMO Academy successfully adapted the DeepSeek-R1 series distillation model on a chip powered by the RISC-V processor Xuantie C920. The entire process took only one hour, demonstrating a swift and seamless experience. This signifies that the DeepSeek series models can be smoothly deployed and run on the full range of Xuantie CPU platforms and other AI end-side devices equipped with RISC-V architecture chips.
RISC-V’s prominence stems from several key factors. Firstly, as an emerging instruction set architecture, it distinguishes itself from the closed or paid licensing models of x86 and ARM by embracing a fully open-source approach. This open-source spirit aligns naturally with the ethos of open-source AI. Its open nature has attracted the participation of over 1,000 companies worldwide, fostering rapid growth in its ecosystem, from hardware design to software toolchains. According to the RISC-V International Foundation, more than 80 different RISC-V chip products have already entered the market.
Secondly, RISC-V offers remarkable flexibility and scalability. It allows developers to customize the instruction set according to specific needs and application requirements. The modular nature of its instruction set enables customization for different application scenarios, a level of flexibility unmatched by traditional, more rigid architectures.
Technically, RISC-V is also exceptionally well-suited for new types of AI computing. Its vector extension (V-extension) can effectively handle large-scale parallel operations, meeting the efficiency demands of AI computation. The open architecture of RISC-V can work in synergy with hardware acceleration modules to enhance the execution efficiency of AI tasks. Through deep integration with AI algorithms, the RISC-V architecture can be used to design dedicated hardware acceleration units, optimizing performance for specific AI models and workloads.
Therefore, many seasoned experts in the chip industry anticipate that RISC-V will become the native computing architecture of the AI era, providing a foundation for innovation and growth.
At the third Xuantie RISC-V Ecosystem Conference hosted by Alibaba’s DAMO Academy, this expectation took a significant step towards becoming a reality.
Xuantie’s First Server-Grade CPU Set for Delivery: A Fusion of High Performance and AI
At the conference, Ni Guangnan, an academician of the Chinese Academy of Engineering, stated, ‘Open-source RISC-V is not only a technological innovation but also a global transformation that will influence the future of computing architecture.’ As a chip instruction set architecture ‘born open-source,’ RISC-V has demonstrated remarkable performance in this semiconductor industry cycle. It has accelerated its progress from embedded systems to complex scenarios such as high-performance computing, offering a new and compelling option for AI computing power.
Among the 25 standards approved by the RISC-V International Foundation in 2024, more than half are related to high performance or AI. Lu Dai, Chairman of the Board of Directors of the RISC-V International Foundation, stated at the conference that one of the most exciting advancements in the RISC-V instruction set is the Matrix extension, which will propel RISC-V to become a formidable force in the AI field.
It is predicted that by 2030, the overall market share of RISC-V will reach 20%, with its share in AI accelerators potentially exceeding 50%. This represents a significant shift in the landscape ofcomputing architecture.
At the conference, DAMO Academy unveiled its next-generation flagship processor, and the first server-grade processor, the C930.
The C930 achieves a general-purpose computing power benchmark of 15/GHz in the SPECint2006 benchmark test. What does this signify? Academician Ni Guangnan pointed out that for RISC-V to truly enter the high-performance computing market, it must achieve a high-performance score exceeding 15 in the SPECint 2006 software test. Therefore, the C930 represents a milestone step for RISC-V, demonstrating its capability to compete in the high-performance arena.
Furthermore, the C930 is equipped with dual engines: 512-bit RVV1.0 and 8 TOPS Matrix. This integrates general-purpose high-performance computing power with AI computing power natively. It also provides an open DSA (Domain-Specific Architecture) extension interface to support more feature requirements, allowing for customization and optimization for specific workloads.
Simultaneously, DAMO Academy disclosed its development plans for new members of the Xuantie processor family, including the C908X, R908A, and XL200, continuing to evolve in directions such as AI acceleration, automotive applications, and high-speed interconnection. Specifically, the C908X is positioned as Xuantie’s first dedicated AI processor, supporting a 4096-bit ultra-long data bit width RVV1.0 vector extension. The R908A is targeted at the high-reliability requirements of automotive-grade chips. The XL200 will provide larger-scale, higher-performance multi-cluster coherent interconnection.
To complement the capabilities of Xuantie processors, DAMO Academy has also launched three Xuantie SDKs based on the three mainstream operating systems: Linux, Android, and RTOS. These SDKs comprehensively integrate Xuantie’s accumulated software capabilities over the years, providing them to the industry in a more complete, convenient, and stable manner. Among them, the Xuantie Linux SDK offers a rich set of subsystems, including Hypervisor virtualization, CoVE (Confidential Virtual Environment) security framework, Xuantie AI framework, and high-performance operator libraries, facilitating the development of RISC-V in high-performance and AI scenarios.
While developing high-performance hardware and software technologies, Xuantie is also actively driving collaborative innovation among upstream and downstream industry partners, accelerating the deployment of a comprehensive RISC-V ‘high-performance + AI’ ecosystem.
Alibaba’s Dedication: RISC-V Xuantie Leads the International Open-Source Community
For those unfamiliar with Xuantie, here’s a brief introduction to its history and significance.
In 2018, Alibaba established the Xuantie brand, focusing on the RISC-V direction. A year later, the first processor, the C910, emerged as the most powerful RISC-V processor at the time. Since then, Xuantie has been a leader in the international RISC-V ecosystem and one of the largest Chinese contributors to the international open-source community. It currently holds chairman or vice-chairman positions in the foundation’s technical committee and more than 10 technical subcommittees, actively promoting the standardization of AI-related technologies.
Since 2019, Xuantie has launched 13 RISC-V processors, covering various scenarios such as high performance, high energy efficiency, and low power consumption. These include:
- C Series (Computing): Primarily targeting high-end servers, high-end edge computing, and industrial/consumer-grade IPCs (Industrial PCs).
- E Series (Embedded): Mainly used in high-end MPUs (Microprocessor Units) and various MCUs (Microcontroller Units).
- R Series (Reliability & Realtime): Targeting high-end SSDs, communications, high-end industrial control, automotive, and other scenarios.
- XT-Link: A CPU multi-cluster interconnect IP (Intellectual Property).
To date, Xuantie processor shipments have exceeded 4 billion units, making it one of the most influential and market-leading processor product series in the domestic RISC-V field.
Throughout its development, Xuantie has consistently pushed the performance boundaries of RISC-V, striving for ever-higher performance. Simultaneously, it has actively embraced AI, aiming to establish RISC-V as a native AI computing architecture.
At the instruction set architecture technology level, leveraging the superior openness and flexibility of the RISC-V architecture, Xuantie has long customized instruction set extensions for AI applications. Its proposed Matrix extension instruction set and optimization of the GEMM (General Matrix Multiply) core operator for large models can accelerate AI inference and training, improving the energy efficiency of AI on edge devices.
In terms of processors, the Xuantie C907 was the first to implement the Matrix extension, achieving a 15x speedup compared to traditional solutions. The upgraded C920 supports Vector 1.0 and Vector Crypto technologies, improving GEMM performance by over 7x and Transformer operator performance by over 17x. The latest flagship processor, the C930, features both vector and matrix dual engines, positioning it as a promising partner for large AI models on edge devices.
At the software stack level, Xuantie has created an end-to-end RISC-V AI full-stack software and hardware platform. This platform provides chip manufacturers with a general-purpose, efficient AI computing infrastructure, forming a pipeline design oriented towards business needs, truly enabling convenient and deep optimization from underlying hardware design to upper-layer software toolchains. This platform has been applied to terminal products such as cloud video transcoding cards, AI edge computing boxes, and RISC-V laptops.
In addition to its own technology, the DAMO Academy RISC-V team has consistently engaged upstream and downstream industry partners to enhance the ‘high-performance + AI’ ecosystem of RISC-V.
At last year’s conference, the RISC-V open-source laptop ‘Ruyi BOOK Jia Chen Edition’ made a surprise appearance, demonstrating stable and smooth operation of large commercial software. This year, the Institute of Software, Chinese Academy of Sciences, further introduced the ‘Ruyi BOOK Yi Si Edition,’ intelligent robots, AI PCs, and other RISC-V high-performance applications.
Among them, the AI PC prototype based on the C920 has successfully run open-source models such as Llama, Qwen, and DeepSeek, supporting AI applications such as AI personal assistants, AI programming, and visual recognition. This demonstrates a complete ‘open-source AI full chain’ from open-source hardware architecture to open-source operating systems and open-source AI models, while also reducing unit computing energy consumption by 30%.
Furthermore, Xuantie has collaborated with partners to build practical solutions such as RISC-V video codec solutions and cloud desktop solutions. To support applications in more industries, Xuantie has also deployed RISC-V computing power in all-in-one PCs, industrial control AI, robots, and other fields.
Academician Ni Guangnan stated that Xuantie’s pragmatic investment and innovation are crucial driving forces for the healthy development of the RISC-V ecosystem.
The Future of Open Source
DeepSeek’s success is a testament to the power and potential of open source. The open-source instruction set architecture RISC-V, since its inception over a decade ago, has charted a different development path from the closed x86 and the licensed ARM models. It has presented the industry with an opportunity to innovate architectures in a more concise and open manner, gaining increasing recognition and adoption.
It is emerging as the best candidate for the native architecture of the AI era. On the one hand, RISC-V, with its commitment to openness and continuous evolution, can keep pace with the rapid changes and advancements in AI. On the other hand, RISC-V’s strong extensibility allows it to be compatible with existing architectural ecosystems through porting and adaptation, while also serving as a native architecture to support emerging scenarios and applications.
As Guo Songliu, head of RISC-V at the Institute of Software, Chinese Academy of Sciences, said: ‘The AI software stack is still evolving rapidly. As the most flexible and open of the three mainstream instruction set architectures, RISC-V is undoubtedly the most suitable for the pace of technological innovation in the AI era.’ The synergy between open-source AI and open-source hardware, exemplified by the collaboration between DeepSeek and RISC-V, promises a future of accelerated innovation and wider accessibility in the world of computing.