AI STORAGE
Designed for massive unstructured data
Deliver high-throughput, low-latency access for petabyte-scale AI training data
Core Capabilities Matrix
Dimensions
- Concurrent Throughput
- IOPS
- Scalability
- Protocol Support
Performance Metrics
- 320 GB/s linear read/write
- 2M random I/O
- 100PB single namespace;
- Supports S3/NFS/CSI
Business Value
- Zero data waiting during thousand-card training
- Sub-second loading of millions of small files
- Eliminates data sharding management
- Seamless TensorFlow/PyTorch integration
Service Benefits
Fast
Providing customized data migration solutions for rapid data transfer across multiple public clouds, enabling quick AI project launches
Scalable
The scalable AI storage solutions support TB to PB-level expansions on demand
Reliability
Highly reliable and observable infrastructure, with full-stack troubleshooting capabilities (e.g. IB networks and servers), ensures rapid resolution of issues
Core Capabilities Matrix
Distributed all-flash array
- Suitable for high-frequency checkpoint saving (periodic full model snapshots)
- 4K random read/write latency < 200 µs.
Tiered Storage Engine
- Hot data: accelerated by an NVMe pool integrated with a data pre-warming service
- Cold data: tiered down to object storage (reducing costs by 70%)
