The surge in Agent technology, driven by models like "OpenClaw" this year, has created significant semiconductor demand, elevating many suppliers. On March 24, Alibaba DAMO Academy unveiled its key chip products—the new flagship CPU Xuantie C950 and the high-efficiency CPU Xuantie C925. These are designed specifically to meet the hardware demands of the rapidly growing AI Agent market. Industry experts note that, based on metrics like clock speed, this marks the first time a RISC-V open-source architecture CPU has achieved single-core performance comparable to high-end ARM or earlier-generation x86 processors, reaching commercially viable server-level single-thread performance. RISC-V has long been perceived as suitable for low-end, edge, and lightweight applications, but the Agentic AI era is changing that. While past AI bottlenecks centered on GPU computing power, the current phase sees constraints spreading to memory, I/O, and system scheduling, shifting the computing paradigm from reliance on GPUs to heterogeneous architectures. Unlike traditional computing chips that prioritize FLOPS, these RISC-V CPUs focus on organizing distributed computing resources, positioning themselves as the operational core of AI systems. Lu Da, Chairman of the RISC-V International Foundation Board, stated that there is strong market demand for major international companies to adopt RISC-V as a primary product. For Alibaba DAMO Academy, years of preparation have culminated in an opportunity to capitalize on this trend.
In the AI Agent era, Alibaba aims to define high-end CPUs to meet emerging demands. The Xuantie C950, built on the open-source RISC-V architecture, leverages its flexibility and customizability, integrating a self-developed AI acceleration engine. It natively supports large-scale models like Qwen3 and DeepSeekV3, which have hundreds of billions of parameters. Unlike proprietary architectures, RISC-V is considered an emerging framework "born for AI," potentially reshaping the chip industry. The C950’s performance, such as surpassing 70 points in the SPECint2006 benchmark with a single-core performance exceeding 22/GHz and a maximum clock speed of 3.2GHz, demonstrates RISC-V’s entry into high-performance and AI computing. This shift could position RISC-V as a competitor to x86 and Arm architectures, moving beyond its role as a low-cost alternative. The C950’s usability, validated through tests with server workloads like MySQL, Redis, Nginx, and OpenSSL, shows it is ready for real-world applications in cloud computing, generative AI, high-end computing, and edge computing. Supporting RVA23.1 standards, it aligns with server, automotive, and AI platforms, crucial for integration into mainstream operating systems and supply chains. Meng Jianyi, Chief Scientist at Alibaba DAMO Academy, emphasized that while RISC-V has penetrated smart devices, automotive, home appliances, and communications, performance limitations and software ecosystem barriers remain. High-performance benchmark products are essential for RISC-V to compete with traditional architectures and capture opportunities in the AI era.
Previously, AI computing was dominated by GPUs, but the Agentic AI era differs. With multiple intelligent agents running simultaneously, factors like token invocation, KV-Cache loading, first-token latency, and task scheduling elevate CPU importance. Meng highlighted that as model capabilities cross thresholds, increasing AI interactions necessitate redesigned CPU architectures. CPUs are evolving from supporting roles to central system coordinators. DAMO Academy introduced two RISC-V native AI computing engines—a 4K ultra-wide Vector engine and a Matrix engine—unified with the CPU to eliminate data copy bottlenecks and integrate general-purpose and AI computing. The C950 smoothly runs top-tier models like Qwen3 and the full-version DeepSeekV3, achieving output speeds of 34 tokens/s and 18 tokens/s, respectively, with low latency. This signifies RISC-V CPUs natively supporting hundred-billion-parameter models, shifting from general-purpose CPUs to AI Agent computing hubs. Alibaba’s RISC-V journey began in 2018, with the 2019 release of the Xuantie C910, one of the industry’s highest-performing RISC-V CPU IPs, breaking the 2GHz barrier. Subsequent developments, including a stable RISC-V laptop and cloud instances, demonstrate Alibaba’s commitment to advancing RISC-V toward commercial, high-performance applications. Meng noted that transforming standards into mass-produced chips requires long-term investment, focusing on ecosystem value over five to ten years.
A true competitive moat involves architecture, ecosystem, standards, and industry collaboration. Alibaba is building an open RISC-V ecosystem with its Flex platform, offering processor modeling, development environments, and software toolchains. This allows customers to use standardized Xuantie CPUs as a base while enabling deep customization. Last year, Xuantie supported 35 clients in 38 CPU modifications, focusing on AI acceleration, storage optimization, and reliability. This platform approach, rather than merely selling IP, lowers barriers to high-end custom chips. Regarding fragmentation concerns, Lu Da stated that RISC-V’s flexibility allows custom extensions, which can evolve into official standards, fostering innovation. Meng added that standards ensure operating system compatibility, while leaving room for innovation drives RISC-V’s growth. For industry clients, this reduces customization costs; for the ecosystem, it enables practical implementation. Huang Shaorui, General Manager of Allwinner Technology’s Product R&D Center, cited growing demand for general-purpose computing in products like smart robots, where multi-core capabilities are essential. Jiang Tao, VP of Resource Development at Southchip Semiconductor, highlighted that without open-source kernel support, power management chips would struggle with digitalization challenges. Though RISC-V is still maturing, its cost competitiveness and modular flexibility help companies develop high-end products. Alibaba’s involvement in RISC-V international standards, such as server-level chips and key specifications, underscores its role in shaping the ecosystem. As Meng stated, by building AI acceleration engines on RISC-V standards like RVV, they avoid the pitfalls of isolated NPUs, fostering long-term ecosystem vitality. This represents a competition for ecosystem influence rather than just products.
In discussions, Meng highlighted that time is the main challenge for RISC-V adoption, with a four-year cycle from standards to mass production. He noted that inference is a key focus in 2024, as AI models require deeper integration, shifting to Agentic AI. CPU architectures must be redesigned for AI interactions, with Alibaba’s C950 optimized for performance, access, and security. Regarding Physical AI, Huang mentioned RISC-V’s edge in edge and endpoint applications due to rapid ecosystem development. Lu Da emphasized the need for international companies to adopt RISC-V as a main product. Meng confirmed that Alibaba’s models use Xuantie for inference, with ongoing adaptations. Xiao Jianhong, CEO of ChipWing Technology, pointed to Shanghai’s advantages in semiconductor innovation and open-source communities, with RISC-V becoming essential for global products. Jiang Tao discussed power management chips' complexity, requiring RISC-V kernels for digital integration. Meng explained that AI accelerators based on RISC-V standards, unlike proprietary NPUs, support ecosystem growth. He acknowledged RISC-V’s head-start effect, requiring significant investment, which Alibaba is committed to. The C950 balances generality and customization through performance benchmarks and instruction set optimizations, enabling market-specific enhancements. Lu Da and Meng agreed that standardization and customization coexist in RISC-V, driving innovation through market evolution. Huang cited Alibaba’s long-term investment and product alignment as reasons for choosing Xuantie, yielding customization benefits and implementation efficiency. Jiang noted RISC-V’s cost competitiveness and flexibility for Southchip’s growth. Meng concluded that while AI trends like OpenClaw influence chip development, the focus is on Agentic AI’s broader demands, requiring strengthened CPUs and optimized CPU-GPU integration for future systems.
Comments