Original Link:Leading 1.5 Generations Ahead, the World’s First 8-Core RISC-V AI CPU Released

Humanity is about to enter a new era of Robot Computers. This new era is driven by artificial intelligence and large models, where vision, language, and AI become the core modes of interaction, representing a proactive interaction style, which differs greatly from the passive interaction methods of personal computers and mobile devices.

As we move from personal computers to smartphones, the demand for computational power has been accelerating. Facing the challenges of large models (LLM) and AIGC, the need for higher computational efficiency, lower power consumption, and reduced costs is becoming increasingly urgent. The growing gap between the demand for and the supply of computational power is becoming more evident.

From observing the iterative development of chips, computing power, and AI, it is clear that high-energy-efficiency computing, a universal software stack, and highly optimized compilers are essential for the success of AI computing products. Recently, SpacemiT made significant strides in high-performance computing chips with the successful launch of the world’s first 8-core RISC-V AI CPU, SpacemiT Key Stone® K1, completing the entire technical and commercial loop for RISC-V high-performance computing.

On April 29th, during the “Creative Core—Endless Innovation” annual product launch event by SpacemiT, Dr. Chen Zhijian, founder and CEO, introduced the K1 chip. He presented real-world test data showing that in terms of AI algorithm model iteration and deployment, the K1 chip, with a comparable microarchitecture, is at least 1.5 generations ahead of ARM chips.

Dr. Chen explained that current mainstream AI algorithms are usually deployed using abstract multi-framework models that are applied to specific chip platforms, such as CPUs, GPUs, and NPUs. In the edge and terminal computing fields, many chip companies lack the ability to customize AI computing power based on CPU cores or licenses. While NPUs have some advantages in power consumption, their use is widespread. However, NPUs have a critical drawback: each NPU has its own unique software stack, creating a closed ecosystem with poor interoperability and integration with other platforms. This lack of transparency makes secondary development (such as deploying private operators) difficult, often requiring collaboration with chip manufacturers, IP vendors, and software stack maintainers.

To address these challenges, SpacemiT adopted an open strategy for its K1 chip design and ecosystem. The K1 is based on a general-purpose CPU with minimal Domain-Specific Architecture (DSA) customization following the RISC-V IME extension framework, along with significant microarchitecture innovations. It maximizes the reuse of open-source resources, providing TOPS-level AI computing power while accelerating edge AI applications. This open compatibility reduces the barriers to deploying existing AI models and allows for more innovative possibilities, transforming AI solutions into a domain that any programmer can contribute to.

“This openness and compatibility not only lower the barriers to deploying a large number of existing AI models but also provide more opportunities for innovation, making AI solutions no longer a specialized field, but an area where every programmer can participate and innovate.” — Dr. Chen Zhijian, Founder and CEO of SpacemiT

Dr. Chen also emphasized that, based on SpacemiT’s AI technology roadmap, the K1 chip supports the rapid deployment of many open-source models through a lightweight plugin system and an open software stack. To date, approximately 150 models, spanning image classification, segmentation, object detection, speech recognition, and natural language understanding, have been optimized for deployment. The K1 chip supports open-source model repositories such as timm, onnx modelzoo, and ppl modelzoo with nearly 100% success, and theoretically, it can support all publicly available ONNX models.

In a further commitment to the open-source community, Dr. Chen announced that JinDie Shikong would release all self-developed AI extension instructions and AI software stack source code to share AI technology with the RISC-V ecosystem and contribute to its growth.

The K1 chip has not only made breakthroughs in AI, but also achieved significant advancements in storage, computation, and floating-point performance, surpassing ARM’s Cortex-A55 in all three areas.

At this year’s annual product launch event, SpacemiT not only unveiled its self-developed SpacemiT Power Stone® P1 power chip and SpacemiT Key Stone® K1 terminal chip but also showcased ecosystem products such as the SpacemiT Muse Book, SpacemiT Muse Shelf, and SpacemiT Muse Box, all powered by the K1 chip. Dr. Chen Zhijian revealed during the event that the SpacemiT Muse series ecosystem products will soon be available for pre-sale, providing more convenient hardware and software product support for the global RISC-V ecosystem.