On March 14, 2024, the 2nd Xuantie RISC-V Ecosystem Conference themed "Open · Connect" was held in Shenzhen. Industry experts, technical leaders, enterprise decision-makers and RISC-V community developers from around the world attended the conference to witness the latest progress of Xuantie RISC-V in various fields such as technological innovation, application expansion and ecological achievements. Axera, an artificial intelligence chip R&D and basic computing power platform company, was invited to participate in the event.

From cloud-side large models represented by ChatGPT, to edge-side computing boxes and various cameras scattered at the end side, AI has become ubiquitous with the rapid development of technology. AI has begun to play a role in a variety of end-side business scenarios such as intelligent driving, intelligent robots, smart factories, smart cities and smart retail, bringing productivity improvements to all walks of life.

As a platform-based chip company supporting edge and end-side intelligence, Axera takes "Democratizing AI for a Better Life" as its corporate mission, and focuses on the R&D of high-performance, low-power artificial intelligence vision processor chips. Liu Jianwei introduced that Axera's series of chip products have been widely applied and achieved steady development in two major product lines of AIoT and ADAS, promoting the deployment and implementation of AI at the edge and end sides.

However, the implementation of AI at the edge and end sides also means higher computing power demand. In this regard, the concise, flexible and efficient RISC-V instruction set architecture has shown its advantages and is expected to become one of the core processor architectures in the era of the Internet of Everything. As a Xuantie Preferred Partner this time, Axera also shared its industrial implementation practices based on Xuantie RISC-V on site. For example, in the IoT field, by applying the Xuantie RISC-V processor and combining it with two self-developed core technologies: Axera Smart Vision AI-ISP and Axera Smart Compute mixed-precision NPU, Axera has realized fast and low-power AI detection and wake-up, which can greatly reduce system power consumption and easily meet the needs of battery-powered products.

At the same time, through the combination of Axera Smart Compute mixed-precision NPU and Xuantie RISC-V IP, Axera has built a high-energy-efficiency AI computing platform, realizing large computing power applications such as image segmentation and text-based image search, and meeting the demands for perception, computing and data processing capabilities in different edge-side scenarios.

Computing power is the foundation of AI. So in the AI era, what kind of computing power infrastructure do we really need?

Liu Jianwei pointed out that at present, the computing density of AI programs has become higher, calling for the emergence of NPU as a new type of processor. "As a dedicated AI processor, NPU should have its own instruction set, rather than being simply understood as an accelerator or an extended instruction set," he said. He indicated that in AI computing, the objects of instruction operations have become high-dimensional tensors, so the energy consumption of the entire processor should be focused on data computing, and the energy consumption of data movement, sorting and combination should be reduced. From the perspective of inference, most AI algorithms/operators have reached a convergent state. Therefore, designing an NPU with operators, the basic computing units of AI models, as the instruction set can open up the hardware microarchitecture design space, allowing more design ideas to be applied and improving energy efficiency ratio.

As Axera's self-developed core technology, Axera Smart Compute mixed-precision NPU is an AI computing processor with operators as the atomic instruction set. It adopts a multi-threaded heterogeneous multi-core design, realizes optimization of operators, network microstructures, data flow and memory access, efficiently supports mixed-precision algorithm design and natively supports Transformer network structure, laying a solid foundation for the application of large models at the edge and end sides.

Meanwhile, Axera Smart Compute mixed-precision NPU also provides a mature and easy-to-use compiler, which greatly reduces the development and deployment costs of AI software, helps customers get started quickly within an hour, and drives the vigorous development of AI computing.

From the perspective of evaluation indicators of AI processors in the visual field, Axera Smart Compute mixed-precision NPU also performs excellently. When running SwinT, its quantization accuracy reaches 80.45%, which is higher than the market average, with high performance of 416 FPS and low power consumption of 199 FPS/W.

To meet the diversified AI computing needs, NPU as an AI processor also requires more flexible dynamic control, which is the reason for Axera's in-depth cooperation with Xuantie RISC-V. It is reported that Axera is exploring the development of a generative AI solution based on Xuantie RISC-V main control and Axera Smart Compute mixed-precision NPU. "We hope to work with Xuantie RISC-V to create the optimal platform choice for AI hardware, bring AI to various industries, and help more customers improve efficiency and create value," said Liu Jianwei.