Appeared at WAIC｜Lingxi Technology LynAInfra (Lingli) brain-like super node makes a grand debut, reconstructing a new pattern of large model reasoning

During the current World Artificial Intelligence Conference WAIC, Lingxi Technology officially released its self-developed brain-inspired super node cluster product - LynAInfra (Lingli). With its full-link self-developed brain-inspired computing power solution, Lingxi Technology solves the pain points of the current AI reasoning industry and brings new options to the domestic high-speed reasoning track. The current development of the artificial intelligence industry is encountering obvious constraints. Two major problems: excessive energy consumption and high inference latency continue to hinder the large-scale implementation of AI inference business. Even if GPU products are iteratively updated, the underlying limitations of traditional chip architectures still cannot be broken; however, the demand for market-side reasoning is exploding, and in the future, reasoning business will account for more than 80% of the AI ​​infrastructure market. With the rise of the token economy, high-value scenarios such as AI programming acceleration, AI video generation, and multi-agent collaboration have put forward higher standards for inference latency and usage costs. The shortcomings of GPU solutions such as high operating expenses and high first token latency have become increasingly prominent. NVIDIA's layout of Groq to increase the number of high-speed inference tracks also confirms that the industry is eagerly looking for inference architecture innovation solutions. Based on the current status of the industry, Lingxi Technology relies on the innate characteristics of self-developed brain-like chips that integrate storage and computing, event-driven, and many-core parallelism, and leverages the core advantages of low power consumption and low latency to create LynAInfra, a high-performance brain-like super node for high-speed reasoning scenarios, benchmarking the innovative version of Groq. Compared with traditional GPU solutions, the product's inference energy efficiency is improved by 3-10 times, and the first token delay is compressed to tens to hundreds of milliseconds. It takes into account powerful inference performance and excellent energy consumption levels, setting a new benchmark for "watts for bits" computing power. LynAInfra Lingli integrated super-node cluster is completely independently developed by Lingxi Technology and is built based on the brain-inspired chips that have achieved large-scale commercialization in China. LynAInfra128 computing power cabinet is equipped with 128 HP640 series boards. Thousands of brain-like chips are cascaded in a single cabinet, releasing nearly 100P@FP16 computing power, and the power consumption of the whole machine is controlled within 30kW. Relying on the self-developed LynxLink high-speed interconnection protocol and unified memory address compilation technology, the product supports both single-cabinet computing power increase (Scale up) and