 
|
NVIDIA ConnectX-8 SuperNIC 800Gb/s PCIe Gen6 Adapter – Single-Port OSFP, InfiniBand/Ethernet, for AI/HPC Clusters
Détails sur le produit:
| Nom de marque: | Mellanox |
| Numéro de modèle: | 900-9x81q-00cn-st0 |
Conditions de paiement et expédition:
| Quantité de commande min: | 1 PIÈCES |
|---|---|
| Prix: | Negotiate |
| Détails d'emballage: | Boîte extérieure |
| Délai de livraison: | Basé sur l'inventaire |
| Conditions de paiement: | T/T |
| Capacité d'approvisionnement: | Fourniture par projet / lot |
|
Détail Infomation |
|||
| Numéro de modèle.: | C8240 (900-9x81q-00cn-st0) | taux de transmission: | 400 gbe |
|---|---|---|---|
| Ports: | Double port | Fonction: | LACP, support empilable, VLAN |
| Technologie: | Infiniband | Interface hôte: | Gen6 x16 |
| Type d'interface: | QSFP112 | Marque déposée: | Millanox |
| Mettre en évidence: | NVIDIA ConnectX-8 SuperNIC 800Gb/s network adapter,PCIe Gen6 OSFP InfiniBand Ethernet adapter,AI HPC clusters network card |
||
Description de produit
NVIDIA ConnectX-8 SuperNIC 800Gb/s PCIe Gen6 Adapter
The industry's highest-performance 800 Gb/s networking adapter, purpose-built for massive-scale AI factories and hyperscale cloud data centers. Featuring PCIe Gen6 host interface and support for both InfiniBand and Ethernet, this SuperNIC delivers extreme throughput, advanced congestion control, and in-network computing acceleration for trillion-parameter AI workloads.
Product Overview
The NVIDIA ConnectX-8 SuperNIC (model: 900-9X81Q-00CN-ST0) represents a generational leap in network connectivity for AI and high-performance computing. Designed as the foundational building block for next-generation AI fabrics, this adapter delivers up to 800 Gb/s of bi-directional throughput, enabling unprecedented scaling for GPU clusters supporting hundreds of thousands of accelerators.
As a key component of the NVIDIA networking portfolio, the ConnectX-8 SuperNIC integrates seamlessly with NVIDIA switches and GPUs, providing end-to-end network performance optimized for both training and inference at scale. Whether deployed in InfiniBand or Ethernet mode, this adapter ensures low latency, high message rates, and power-efficient operation—critical for sustainable AI data center growth.
Key Features
- 800 Gb/s Bandwidth: Supports up to 800 Gb/s total throughput across InfiniBand and Ethernet protocols, with flexible port configurations including 1x OSFP or 2x QSFP112
- PCIe Gen6 Host Interface: Up to 48 lanes of PCIe Gen6 connectivity, delivering massive bandwidth to the host system with support for NVIDIA Multi-Host technology (up to four hosts)
- Advanced Congestion Control: Programmable telemetry-based congestion management ensures predictable performance under heavy AI traffic loads
- In-Network Computing: Full support for NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARP) and GPUDirect RDMA, offloading collective operations from the CPU
- Comprehensive Security: Secure boot with hardware root of trust, flash encryption, device attestation (SPDM 1.1), and inline crypto accelerations for IPsec, MACsec, and PSP
- Precision Timing: Advanced PTP (IEEE 1588v2, G.8273.2 Class C), SyncE, and configurable PPS for time-sensitive applications
Technology Deep Dive
The ConnectX-8 SuperNIC is engineered around a purpose-built ASIC that integrates a 800 Gb/s network interface with a PCIe Gen6 host interface. This architecture enables RDMA and RoCEv2 accelerations, high-speed packet reordering, and burst-buffer offloads that are essential for distributed AI training. The adapter also supports NVIDIA GPUDirect RDMA, allowing direct data movement between GPUs across the network without CPU intervention, dramatically reducing latency and improving application scalability.
For cloud and enterprise environments, the ConnectX-8 SuperNIC includes Accelerated Switching and Packet Processing (ASAP2), providing hardware-accelerated OVS, overlay network offloads (VXLAN, GENEVE, NVGRE), connection tracking, and hierarchical QoS. These features make it equally suitable for virtualized workloads, SDN deployments, and containerized microservices architectures.
Typical Deployments
- AI Training Clusters: Large-scale GPU fabrics utilizing InfiniBand for lossless, low-latency connectivity between thousands of nodes
- AI Inference Farms: Ethernet-based deployments supporting real-time inference workloads with RoCEv2 for efficient GPU-to-GPU communication
- Cloud Data Centers: Hyperscale environments requiring high-bandwidth, secure, and programmable network adapters for multi-tenant services
- HPC Research Facilities: Institutions running MPI-based simulations and scientific computing applications that benefit from in-network computing and SHARP offloads
- Enterprise AI Platforms: On-premises AI infrastructure for finance, healthcare, and manufacturing sectors requiring data sovereignty and predictable performance
Compatibility
The ConnectX-8 SuperNIC is fully compatible with:
- NVIDIA Quantum-2 and Spectrum-4 Switches: End-to-end NVIDIA networking solutions for InfiniBand and Ethernet fabrics
- NVIDIA HGX and DGX Platforms: Seamless integration with NVIDIA GPU server architectures
- Major Operating Systems: Linux (all major distributions), Windows Server, and VMware ESXi
- AI Software Stack: NVIDIA NCCL, HPC-X, DOCA UCC/UCCX, Open MPI, and MVAPICH2
- Industry Standards: PCIe Gen6, InfiniBand Trade Association (IBTA) v1.7, IEEE 802.3 Ethernet standards
Note: For specific chassis and server compatibility, please verify mechanical dimensions and thermal requirements before ordering.
Technical Specifications
| Parameter | Specification |
|---|---|
| Product Model | 900-9X81Q-00CN-ST0 |
| Network Protocols | InfiniBand and Ethernet |
| Maximum Total Bandwidth | 800 Gb/s |
| InfiniBand Speeds | 800 / 400 / 200 / 100 Gb/s |
| Ethernet Speeds | 400 / 200 / 100 / 50 / 25 Gb/s |
| Host Interface | PCIe Gen6 (up to 48 lanes) |
| Form Factor Options | PCIe HHHL 1P × OSFP, HHHL 2P × QSFP112, Dual ConnectX-8 Mezzanine, OCP 3.0 TSFF 1P × OSFP |
| Port Configuration | 1×800/400/200/100 Gb/s or 2×400/200/100 Gb/s; supports up to 8 split ports |
| RDMA Support | RoCEv2, GPUDirect RDMA, GPUDirect Storage |
| In-Network Computing | SHARP, collective operations, MPI accelerations |
| Security Features | Secure boot, flash encryption, SPDM 1.1, IPsec/MACsec/PSP crypto |
| Timing & Synchronization | PTP (IEEE 1588v2, G.8273.2 Class C), SyncE, PTM, PPS in/out |
| Management Interfaces | NC-SI, MCTP, PLDM (DSP0248, DSP0267, DSP0218), SPDM, SPI, JTAG |
| Network Boot | InfiniBand / Ethernet PXE, iSCSI, UEFI |
| Power Efficiency | Optimized for sustainable AI data center operations |
Selection Guide
| Ordering Option | Description | Use Case |
|---|---|---|
| 900-9X81Q-00CN-ST0 | ConnectX-8 SuperNIC, PCIe HHHL 1P × OSFP | Single-port 800G deployments, standard rack servers |
| HHHL 2P × QSFP112 | Dual-port QSFP112 configuration | 400G per port, active-active failover, high availability |
| Dual ConnectX-8 Mezzanine | Mezzanine module for dense GPU servers | DGX/HGX platform integration, space-constrained designs |
| OCP 3.0 TSFF 1P × OSFP | OCP 3.0 form factor with single OSFP port | Open Compute Project deployments, hyperscale data centers |
Why Choose the ConnectX-8 SuperNIC
- Unmatched AI Performance: Purpose-built for trillion-parameter models, delivering the lowest latency and highest message rates for distributed training
- Future-Ready Architecture: PCIe Gen6 support and 800 Gb/s bandwidth ensure a clear upgrade path for next-generation GPU platforms
- Operational Efficiency: Advanced telemetry and congestion control reduce tail latency and maximize effective AI workload throughput
- Software Ecosystem: Full NVIDIA DOCA, NCCL, and HPC-X support enables rapid development and deployment of AI applications
- Sustainability: Power-optimized design contributes to lower PUE and reduced operational costs at scale
Service & Support
Hong Kong Starsurge Group Co., Limited provides comprehensive pre-sales and post-sales support for the NVIDIA ConnectX-8 SuperNIC portfolio:
- Technical Consultation: Architecture design and optimization for AI fabrics
- Integration Services: System integration with NVIDIA switches, GPUs, and server platforms
- Warranty & RMA: Standard manufacturer warranty with expedited replacement options
- Global Delivery: Worldwide shipping with secure packaging and customs clearance support
- Multilingual Support: English, Mandarin, and Cantonese technical and sales assistance
Frequently Asked Questions
What is the maximum bandwidth supported by this adapter?
The ConnectX-8 SuperNIC supports up to 800 Gb/s total bandwidth in both InfiniBand and Ethernet modes.
Is this adapter compatible with PCIe Gen5 servers?
Yes, it is fully backward compatible with PCIe Gen5, Gen4, and Gen3 slots, though performance will be limited to the host interface capability.
What operating systems are supported?
Major Linux distributions, Windows Server, and VMware ESXi are officially supported.
Does this adapter support RoCEv2?
Yes, full RoCEv2 acceleration is included, with advanced congestion control and GPUDirect RDMA support.
Can I use this adapter in a virtualized environment?
Yes, with SR-IOV, ASAP2, and overlay network offloads (VXLAN, GENEVE, NVGRE), it is ideal for SDN and NFV deployments.
Precautions
- Ensure adequate server airflow and cooling to maintain thermal specifications; avoid blocking the adapter's heatsink area
- Use only certified cables and transceivers for 800 Gb/s operation; refer to the NVIDIA compatibility list
- Verify physical clearance for full-height or low-profile brackets; some servers may require specific form factors
- Firmware updates are recommended for optimal performance and security; use the NVIDIA firmware update tool
- For ESXi deployments, ensure the host server supports PCIe pass-through and SR-IOV for virtual function assignment
About Hong Kong Starsurge Group Co., Limited
Hong Kong Starsurge Group Co., Limited is a technology-driven provider of network hardware, IT services, and system integration solutions. Founded in 2008, the company serves customers worldwide with products including network switches, NICs, wireless access points, controllers, cables, and related networking equipment.
Backed by an experienced sales and technical team, Starsurge supports industries such as government, healthcare, manufacturing, education, finance, and enterprise. The company also offers IoT solutions, network management systems, custom software development, multilingual support, and global delivery. With a customer-first approach, Starsurge focuses on reliable quality, responsive service, and tailored solutions that help clients build efficient, scalable, and dependable network infrastructure.
Key Facts at a Glance
800
Gb/s Max Bandwidth
PCIe 6.0
Host Interface
48
PCIe Lanes
InfiniBand + Ethernet
Protocol Support
SHARP
In-Network Computing
Compatibility Matrix
| Component | Compatibility Status | Notes |
|---|---|---|
| NVIDIA Quantum-2 Switches | Fully Supported | InfiniBand NDR fabrics |
| NVIDIA Spectrum-4 Switches | Fully Supported | 400G Ethernet fabrics with RoCEv2 |
| NVIDIA HGX/DGX Platforms | Verified | Mezzanine form factor required for some |
| Third-party PCIe Gen6 Servers | Varies by Vendor | Check host BIOS and slot power limits |
| Linux (RHEL, Ubuntu, SLES) | Certified | NVIDIA OFED driver support |
| Windows Server | Supported | Check driver version for specific builds |
| VMware ESXi | Supported | Native and passthrough modes available |
Buyer Checklist
- Form Factor: Confirm your server supports PCIe HHHL, OCP 3.0, or mezzanine form factor
- Slot Type: Verify the available PCIe slot supports Gen6 (or Gen5/Gen4) and provides sufficient power (up to 48 lanes)
- Port Configuration: Choose between single-port OSFP or dual-port QSFP112 based on your cabling strategy
- Network Protocol: Select InfiniBand for lossless AI fabrics or Ethernet for multi-protocol environments
- Cooling and Airflow: Ensure your chassis provides adequate airflow over the heatsink; consider active cooling for high-density configurations
- Software Stack: Check NCCL, HPC-X, or DOCA version compatibility with your AI framework
- Warranty and Support: Confirm the warranty period and whether extended support is required for your deployment
Related Products
Vous voulez en savoir plus sur ce produit