Aliyun VS Baidu Cloud

The following is a detailed comparison of computing power: - **Computing power scale**:

** Ali Cloud **: It has Zhangbei Super Intelligent Computing Center and Ulanqab Intelligent Computing Center, among which the total construction scale of Zhangbei Super Intelligent Computing Center is 12000PFLOPS, and the construction scale of Ulanqab Intelligent Computing Center is 3000PFLOPS.

**Baidu Cloud**: There are Baidu Intelligent Cloud-Kunlun Core (Yancheng) Intelligent Computing Center and Baidu Intelligent Cloud (Jinan) Intelligent Computing Center, etc. The computing power of Yancheng intelligent computing center reaches 200PFLOPS, while the specific computing power of Jinan intelligent computing center is not mentioned

**Hardware and Architecture**

**Alibaba Cloud**: Centered around NVIDIA H100, the system achieves full connectivity through NVLink 4.0 with 8 single-chip cards and 4 Mellanox CX7200Gbps network cards. Utilizing a three-layer non-blocking Fat-Tree topology, it supports linear expansion of up to 1024 GPUs per Pod. The Leaf switch is equipped with 128400Gbps ports, delivering a total bandwidth of 51.2Tbps.

**Baidu Cloud**: Through its integrated hardware-software technology system combining "Kunlun Chip (China-made) + BaiGe GPU Computing Platform", Kunlun Super Node single cabinets can accommodate 32/64 Kunlun AI accelerator cards. The fully interconnected communication between cards within a single cabinet achieves bandwidth enhancement of up to 8 times, with the computing power of one cabinet reaching the level of 8 traditional 8-card servers.

Computing power performance

** Ali Cloud **: In the training practice of Tongyi Qianwen, the 235B parameter MoE model only takes 14 days to complete the training on the H100 cluster with 12,000 cards, and the MFU reaches 68%, which reduces the cost by 53% compared with the traditional architecture.

**Baidu Cloud**: Based on the GPU computing power platform of Bai Ge and Kunlun Chip P800, it provides services for open source models such as DeepSeek. The throughput performance of a single card is 90% higher than that of domestic mainstream chip solutions, and the model reasoning speed can be improved by more than 40% in large-scale and high-concurrency scenarios.

**Application scenario adaptation**

**Ali Cloud**: It is famous for its elastic container instance (ECI) and Shenlong architecture. It performs well in large-scale parallel computing, especially suitable for latency-sensitive application scenarios, and fully supports deep learning frameworks such as TensorFlow and PyTorch.

**Baidu Cloud**: Through the optimization of its PaddlePaddle framework, it improves the efficiency of specific AI tasks and has unique competitiveness in application scenarios based on PaddlePaddle framework.