Core Technologies

Based on the foundation of proprietary technologies and a core “brain” built on a deep learning platform, SenseTime has rapidly opened up AI application in multiple vertical scenarios.

Deep Learning Platform

Deep Learning Platform

Technical Capabilities

AI Supercomputing Platform
Self-developed AI Training Framework
High-Performance AI Storage
High-Performance Heterogeneous Computing

AI Supercomputing Platform

The self-developed, large-scale AI supercomputing platform can run 14, 000+ GPUs. With multiple storage backends, it conducts optimizations specific to AI application characteristics, supports storage for hundreds of billions of files, and reads over 1 million files per second. With high performance network and lightweight virtualization, it provides strong computing power.

Self-developed AI Training Framework

SenseTime has developed the AI training framework SenseParrots which owns the entire intellectual property rights for the full technology stack, and supports large-scale concurrent training. It is optimized for industry-scale training, including training in parallel on thousands of GPUs, models with hundreds of billions of parameters, tasks containing tens of millions of categories and tens of billions of samples. Significantly enhancing the efficiency of AI development and deployment, SenseParrots’ overall performance leads the industry and broadens the application boundary.

High-Performance AI Storage

Specific to AI application characteristics, the technology optimizes data transmission, I/O and caching for AI applications. It supports storage for hundreds of billions of files and read performance at over 1 million IOPS. It also simultaneously provides file system and object storage for different data types, guaranteeing data security and system stability with enhanced data isolation, authority control and user-defined QoS. It provides SDKs for users to rapidly perform integration development.

High-Performance Heterogeneous Computing

The advance in modern AI technologies eases the development in computer vision and deep learning technologies with enriched HPC ecosystems. With support of X86, GPUs, ARM, FPGA and other ASICs with different architectures, work can be deployed on HPC data centers, personal computers, mobile platforms and IoT devices. Through the performance optimization, such as automatic code generation and highly efficient scheduling, for different computing architectures, the platform outperforms other competitors by 50% to 200%, providing comprehensive reliability for all kinds of AI applications.