High-spec GPU Supercomputers
NVIDIA‑native DGX architecture and highly scalable inter‑node communication
- DGX H200 is designed end‑to‑end by NVIDIA—from ultra‑high‑speed GPU‑to‑GPU interconnects to storage, networking, power, and cooling—delivering bottleneck‑free performance.It achieves faster computation even compared with hyperscaler environments.
- y adopting RoCEv2 as the communication protocol, the system provides greater scalability and security than InfiniBand.
- Each node is connected via a 1.6 Tbps lossless network, leveraging GDS and RDMA (RoCEv2).
High‑speed connectivity with high‑performance storage
- DDN(30TB)：ombines high performance, large capacity, high reliability, and ultra‑high density, eliminating the most critical I/O bottlenecks in cluster systems(100 Gbps × 8 ports).
- NetApp(20TB)：Delivers speed, efficiency, versatility, and data protection simultaneously(100 Gbps × 8 ports).
High security through logical isolation of tenants and networks
- Logical isolation at the virtual tenant level with access controls, combined with tenant‑level storage encryption.
- Adoption of IP CLOS and Ethernet, with overlay networks used to logically segment multiple networks.
