China TravelSky Cloud Data Partners with Huawei to Build AI Data Infrastructure
Set a new benchmark in office efficiency and AI services.
Set a new benchmark in office efficiency and AI services.
As digital transformation reshapes the aviation industry, China TravelSky Cloud Data Co., Ltd., a subsidiary of China TravelSky Holding Company, has focused its efforts on a core strategic aim—the group's digital and intelligent upgrade. It was looking to build an artificial intelligence (AI) computing platform, powered by a cluster of thousands of neural processing units (NPUs). The platform aims to serve as an intelligent engine to drive service transformation, enhance office efficiency and external service capabilities, and set a new benchmark for intelligent computing upgrade for civil aviation.
Intelligent assistant: Knowledge Q&A for intelligent assistants includes massive amount of files for employees' daily Q&A and compliance checks. Input sequence length exceeded 16k tokens, with insufficient inference throughput to meet concurrent service demands.
Document writing: When drafting analysis reports or service documents, the large AI model generated content based on historical materials—such as civil aviation service cases and policy documents over the past three years—and token counts reached 64k–128k. The input sequence exceeded the context window, resulting in incomplete content, poor logical coherence, and a time to first token (TTFT) of over 30 seconds.
Code generation: For code comprehension using large AI models, the inference sequence length exceeded 8k tokens, with extensive recomputation during processing. The low utilization of computing power and insufficient throughput led to request queuing, failing to meet research and development (R&D) efficiency requirements.
In response to these challenges, TravelSky Cloud Data has now deployed Huawei's AI Inference Acceleration Solution, with OceanStor AI storage system serving as the data backbone of this AI computing platform. Featuring ultimate performance and Unified Cache Manager (UCM-based) inference acceleration, the solution ensures robust data support for civil aviation applications while supporting improvements in operational efficiency and service quality.
Persistent KV cache overcomes the bottleneck of long-sequence inference
The UCM multiplies the inference sequence length, significantly reducing the TTFT and improving inference throughput, making it smooth and fast. This accelerates DeepSeek's analysis and Q&A response, delivering a seamless user experience.
Superb hybrid load performance for faster AI inference
Powered by the innovative scale-out file system OceanFS, data and control plane separation, and NPU Direct Storage (NDS), the Solution delivers TB-level bandwidth and high input/output operations per second (IOPS) performance. This efficiently supplies data for models, enhancing office collaboration and external service response speeds in civil aviation.
A one-stop AI platform solution for easy, efficient application of AI
The ModelEngine AI toolchain and Datacenter Virtualization Solution (DCS) resource scheduling capabilities enable full-stack software and hardware collaboration. The integration of multi-modal data processing, long-sequence inference acceleration, xPU pooling and scheduling, thousands-of-NPU cluster expansion, and full-stack software and hardware O&M eliminates technical barriers, helping TravelSky Cloud Data rapidly build an AI computing platform tailored to its particular service requirements.
Huawei OceanStor AI storage uses innovative technologies to overcome AI inference bottlenecks and efficiently build a thousands-of-NPU AI computing platform. This is empowering the digital and intelligent transformation of China's civil aviation sector, setting a new benchmark for its digital and intelligent upgrade.
