Zhong Jin Zai Xian·2025-08-11 07:53Core Insights - JD Cloud's JoyScale AI computing platform has been upgraded to support the most diverse domestic heterogeneous computing power scheduling in the industry, accommodating over 10 domestic AI computing cards and more than 20 training and inference frameworks, making it the only platform that supports remote invocation of both NVIDIA GPUs and Ascend NPUs [1][3]. Group 1: AI Native Computing Platform - The deep application of AI has created new technical challenges for infrastructure, necessitating an AI Native computing platform that is GPU-centric rather than CPU-centric [2]. - The increasing demand for inference and the need for efficient resource allocation are driving enterprises to rethink their investment in computing resources [2]. Group 2: JoyScale AI Platform Features - JoyScale AI platform is based on JD's internal unified GPU pooling practices, allowing for unified scheduling and resource sharing for training tasks and inference services [3]. - The platform boasts four core advantages: - Extreme computing performance with a 50% improvement in overall inference performance [5]. - Efficient heterogeneous computing power scheduling, achieving a 70% increase in overall resource utilization [5]. - Deep collaboration with domestic AI chip manufacturers to enhance the ecosystem [6]. - Support for over 20 AI training and inference frameworks, including PyTorch, TensorFlow, and Triton [7][8]. Group 3: Performance Optimization - JoyScale has overcome technical challenges for running mainstream models on domestic cards, achieving application and computing power separation for flexible resource allocation [9]. - The platform employs advanced scheduling algorithms to maximize task execution efficiency by optimizing resource allocation based on CPU NUMA and network topology [9]. - In model optimization, techniques such as GE graph compilation and ATB high-performance operator technology have been implemented to enhance inference speed in multi-modal scenarios [10].