A National Joint-Stock Bank AI Computing Power Integration

Solution

Heterogeneous computing power unified scheduling: Through the cloud platform, uniformly manage computing power in multiple data centers, achieving integration and intelligent scheduling of CPU/GPU/NPU multi-type computing resources, supporting real-time (transaction risk control, intelligent customer service) and non-real-time (model training, batch inference) computing power collaboration.
Large model training and inference optimization: Support the bank in introducing open-source large models such as DeepSeek and Qwen, provide full-process services from model evaluation, training tuning to inference deployment, and explore model distillation to reduce inference costs.
Multi-location multi-center intelligent computing network: Build an intelligent computing network covering three locations and five centers in Shanghai, Hefei, and Horinger, implementing the "East Data West Computing" strategy, achieving green and low-carbon operation and optimized allocation of computing resources.
Financial-grade security compliance: Full-stack localization solution meets the bank's stringent data security and compliance requirements, with model training and inference completed within the client's exclusive resource pool, ensuring data does not leave the domain.

Implementation Results

Built a domestic leading thousand-card level localization intelligent computing cluster, supporting AI application landing for more than 10 key projects including inclusive finance, green finance, corporate business, retail innovation, and financial markets.
After unified scheduling of heterogeneous computing power, computing efficiency significantly improved, and model training costs decreased by about 30%.
The intelligent customer service assistant has accumulated more than 300,000 services, and the intelligent risk control model recognition accuracy has improved to over 96%.
Data center planned for 14,000 cabinets, PUE optimized to below 1.2, saving more than 20 million kWh of electricity annually.
Achieved collaborative deployment of large-parameter models, small-parameter models, and multi-modal models, with comprehensive inference costs decreased by 40%.

Customer Testimonial

Chinasoft International helped us build an autonomous, controllable, full-stack localization intelligent computing foundation, forming a complete closed loop from computing power scheduling to model management, providing solid support for the bank's large-scale AI application.