速递｜蚂蚁集团突破技术封锁：国产芯片助力AI训练成本直降20%，性能媲美英伟达H800

Core Viewpoint - Ant Group is leveraging domestically produced semiconductor technology to develop AI model training solutions, aiming to reduce costs by 20% [1][2]. Group 1: Technological Developments - Ant Group is collaborating with Alibaba and Huawei to advance a mixed expert (MoE) machine learning model training based on domestic chips, achieving performance comparable to NVIDIA's H800 chip [1]. - The company has shifted its AI development focus from NVIDIA products to alternatives including AMD and domestic chips [1]. - Ant Group's research paper claims that its model has surpassed Meta Platforms in certain benchmark tests, potentially injecting new momentum into China's AI development [1][4]. Group 2: Cost Efficiency - The cost of training a model with high-performance hardware for 1 trillion tokens is approximately 6.35 million RMB (880,000 USD), which can be reduced to 5.1 million RMB through optimization using lower-spec hardware [3]. - The goal of Ant Group's research is to expand models without relying on high-end GPUs, contrasting with NVIDIA's strategy of developing more powerful chips [2]. Group 3: Model Performance - Ant Group's latest language models, Ling-Plus and Ling-Lite, are designed to provide AI solutions in healthcare and finance, with Ling-Lite outperforming one of Meta's models in English comprehension tests [4]. - Ling-Lite contains 16.8 billion adjustable parameters, while Ling-Plus has 290 billion parameters, making them relatively large in the language model field [4]. - Despite advancements, Ant Group faces challenges in model stability, as minor changes in hardware or model structure can significantly increase error rates [4].