TMTPOST--Axera Technology, a developer of computer vision processors, has launched the “Axera Tongyuan AI Processor” designed for AI edge model inference at the 2024 World Artificial Intelligence Conference (WAIC 2024).

The Axera Tongyuan AI Processor, unveiled last Friday, features an operator instruction set and data flow microarchitecture at its core, using a programmable data flow microarchitecture. It offers three levels of computing power and supports native transformer models.

It has achieved large-scale production in the fields of smart cities and assisted driving and can be applied to various model scenarios such as text-to-image search, general detection, text generation, and AI agents.

Qiu Xiaoxin, the founder and chairman of Axera Technology, noted that the true large-scale implementation of large models requires a tight integration of cloud-edge-terminal systems. The key to the integration of edge and terminal lies in AI computing and perception.

Axera Technology, by leveraging its self-developed core technologies of Axera Zhimou AI-ISP and Axera Tongyuan hybrid precision NPU, has established a strategic route focusing on "AIoT+ADAS" and is advancing into the edge computing and AI inference fields. This aims to accelerate the application of smart cities and intelligent driving.

Qiu pointed out that the team is closely monitoring industry peers' layouts for edge AI model applications. She asserted that the Scaling Law for large models is not a “hard and fast rule” and won’t continue to grow linearly. There will inevitably be periods of stable and gradual optimization, she added.

“People have realized that Moore’s Law has reached its limit. From a single-chip perspective, Moore's Law is still evolving but at a slowing pace. Currently, the semiconductor industry enhances overall performance through system-level solutions. This year’s GTC AI Conference highlighted system-level optimization, not just single chips. The entire system’s scheduling, optimization, and management are critical. The future trajectory of Scaling Law will involve further optimization in some form. Nothing can grow linearly forever; it will reach a point of nonlinear growth. When that inflection point arrives, whether the current optimization methods remain viable needs reevaluation,” Qiu elaborated.

Axera Technology, founded in May 2019, provides foundational AI computing platforms for various industries. The company has successfully developed and mass-produced high, medium, and low-end edge AI chips, focusing on smart cities, intelligent driving, robotics, and innovative business scenarios.

As of now, Axera Technology has completed its Series B financing, with investors including Tencent, Qiming Venture Partners, Meituan, V Hall Venture Capital, Lenovo Capital, and Glory Ventures.

The newly unveiled “Axera Tongyuan AI Processor” is primarily aimed at AI model inference. It optimizes computing power through model algorithm and chip design collaboration and model miniaturization, accelerating the large-scale application of large models.

Qiu said that intelligent computing centers, which other AI chip companies are focusing on, are not Axera’s priority at the current stage. The focus now is on the vast market of edge and terminal applications. According to Qiu, the large-scale application of AI models must involve a cloud-edge-terminal triad.

“The foundational large models of generative AI will definitely be cloud-based. However, whether these large models can be refined or optimized to become industry-specific models and move to the edge, instead of having trillions of parameters, is still possible,” said Qiu. The current stage of large model application is still very primitive, or “broke force,” she added.

Qiu pointed out that the first scenarios for deploying large models at the edge are likely to be in vehicles, followed by smartphones and AI PCs, because vehicles require real-time responses, making them an ideal application for edge large models, such as in intelligent driving, smart cockpits, human-machine interaction, and AI agents.

“Application scenarios are very diverse. A 3.2T small chip integrated into a phone chip can enable many local applications without needing cloud support,” Qiu suggested. She also envisioned that a potential future application for generative AI could be “smart homes,” where a home AI server hub acts as a computing center, with entry points possibly being phones, robots, and "embodied intelligence," camera, and voice control devices.

Regarding the business model for implementing Axera’s AI capabilities, Qiu mentioned two main approaches: One is for clients whose SoC computing power is insufficient or cannot natively support large models, integrating Axera’s NPU IP into their chips to provide efficient NPU capabilities. The other is to achieve large-scale deployment solutions through Axera’s chips and software stack.

Qiu emphasized that the semiconductor industry has long cycles and requires a mutual selection process with investors. Fast-paced, short-term investment firms are unsuitable for this sector.

“Choosing the right investors is crucial. First, investors must understand the entire logic and cycle of the semiconductor industry. Second, as a chip company, achieving a commercial closed-loop is essential. This is why we aim for large-scale production and ecosystem formation. A healthy chip company can achieve breakeven and profitability within seven to ten years on average. We hope to develop steadily and quickly enter a positive cycle,” Qiu remarked.