Workflow
国泰海通|电子:曙光scaleX640重磅发布,国产算力加速突破

Core Viewpoint - The launch of the scaleX640 by Dawning Technology is expected to enhance the performance of trillion-parameter model training and inference by 30-40%, potentially accelerating the breakthrough of domestic computing chips in training scenarios [2][4]. Performance Enhancement - The scaleX640 achieves a 20-fold increase in computing density compared to traditional solutions, enabling a 30-40% performance improvement in MoE trillion-parameter model training and inference scenarios [2]. - The system has undergone a 30-day reliability test, ensuring support for large-scale cluster deployments of up to 100,000 cards, which helps mitigate performance degradation and cost increases during model training due to chip failures [2]. Open Ecosystem - The scaleX640 adopts an AI computing open architecture, supporting multiple brands of acceleration cards at the hardware level and being compatible with mainstream computing ecosystems at the software level [2]. - The open architecture is expected to facilitate the integration of domestic computing software ecosystems and unify supernode structures, potentially aiding rapid iterations of domestic supernodes [2]. Technological Innovation - The scaleX640 utilizes innovative technologies such as ultra-high-speed orthogonal architecture, ultra-high-density blades, immersion phase change cooling, and high-voltage direct current power supply, which may drive technological advancements in the upstream supply chain [3]. Catalysts for Growth - The successful implementation of the scaleX640 in domestic large model training scenarios is anticipated to accelerate the construction of local intelligent computing centers [4].