NVIDIA ConnectX-8 SuperNIC C8180-DT0 800G High-Density AI Fabric
Model: 900-9X81E-00EX-DT0 | Multi-Host | High Density | AI Factory Optimized
The DT0 (Density / Hyperscale) variant is engineered for AI factories operating 10,000+ GPUs. Features NVIDIA Multi-Host technology (up to 4 hosts per adapter), advanced power management (under 28W typical), and telemetry-based congestion control for predictable large-scale performance.
Product Overview - Hyperscale Edition
NVIDIA ConnectX-8 SuperNIC C8180-DT (900-9X81E-00EX-DT0) is purpose-built for hyperscale AI operators. Unlike standard enterprise SKUs, the DT0 variant includes Multi-Host PCIe capabilities allowing a single adapter to serve up to four independent host processors — reducing NIC count per rack and lowering total cost of ownership. Enhanced telemetry and programmable congestion control ensure consistent tail latency across thousands of GPU nodes.
Hyperscale Advantages
- 4:1 Multi-Host consolidation (4 hosts per NIC)
- <28W typical power at full 800G load
- Advanced telemetry: per-flow latency, queue depth, ECN marking
- Adaptive routing support for load-balanced fabrics
- SHARPv3 in-network reduction for collective operations
Scalability Metrics
- Supports AI clusters of 100,000+ GPUs
- 16 million I/O channels per adapter
- Up to 8 logical ports via splitter cables
- Zero packet loss at line rate (PFC + ECN)
- RDMA over Converged Ethernet v2 (RoCEv2)
Multi-Host Architecture
| Configuration |
Hosts Supported |
Total Bandwidth per Host |
Use Case |
| 1 Host |
1 |
800Gb/s |
Maximum single-host throughput |
| 2 Hosts |
2 |
400Gb/s each |
Balanced dual-node density |
| 4 Hosts |
4 |
200Gb/s each |
Maximum rack density |
Target Deployments - Hyperscale AI Factories
- Large language model training clusters (GPT-5 scale)
- Autonomous vehicle training infrastructure
- Cloud provider GPU-as-a-service platforms
- National supercomputing centers
- AI semiconductor design simulation farms
Power & Thermal Optimization
The DT0 variant features enhanced power gating and dynamic voltage scaling, achieving 25-28W typical at 800Gb/s line rate — 15% more efficient than previous generation. Passive cooling sufficient for 1U servers with adequate front-to-back airflow.
