 
|
NVIDIA ConnectX-7 MCX755106AS-HEAT Dual-Port 400Gb/s NDR InfiniBand & 400GbE Smart Adapter Card
제품 상세 정보:
| 브랜드 이름: | Mellanox |
| 모델 번호: | MCX755106AS-HEAT (900-9X7AH-0078-DTZ) |
| 문서: | Connectx-7 infiniband.pdf |
결제 및 배송 조건:
| 최소 주문 수량: | 1PCS |
|---|---|
| 가격: | Negotiate |
| 포장 세부 사항: | 외부 상자 |
| 배달 시간: | 목록을 기반으로 |
| 지불 조건: | 티/티 |
| 공급 능력: | 프로젝트/배치 별 공급 |
|
상세 정보 |
|||
| 모델 번호.: | MCX755106AS-HEAT (900-9X7AH-0078-DTZ) | 포트: | 2 포트 |
|---|---|---|---|
| 기술: | 인피니본 | 인터페이스 유형: | OSFP56 |
| 사양: | 16.7cm x 6.9cm | 기원: | 인도 / 이스라엘 / 중국 |
| 전송 속도: | 200gbe | 호스트 인터페이스: | gen3 x16 |
| 강조하다: | NVIDIA ConnectX-7 400Gb/s InfiniBand adapter,dual-port 400GbE network card,NDR InfiniBand smart adapter card |
||
제품 설명
NVIDIA ConnectX-7 MCX755106AS-HEAT Dual-Port 400Gb/s NDR InfiniBand & 400GbE Smart Adapter Card
NVIDIA Original ConnectX-7 NDR 400Gb/s Dual-Port InfiniBand + Ethernet PCIe Gen 5.0
MPN: MCX755106AS-HEAT
The NVIDIA ConnectX-7 MCX755106AS-HEAT is a flagship dual-port smart network adapter delivering up to 400Gb/s per port (800Gb/s aggregate) for both InfiniBand (NDR) and Ethernet (400GbE) networks. Featuring a dual-port QSFP configuration, PCIe Gen 5.0 x16 host interface, and comprehensive hardware acceleration engines, this adapter is purpose-built for AI factories, scientific computing clusters, hyperscale data centers, and security-sensitive enterprise infrastructures requiring maximum density, ultra-low latency, and extreme throughput.
400Gb/s
Per port speed
NDR InfiniBand / 400GbE
Protocol support
800Gb/s
Aggregate bandwidth
PCIe 5.0 x16
128GB/s bidirectional
Hardware Security
IPsec/TLS/MACsec encryption
IEEE 1588v2
12ns accuracy
Product Overview
The NVIDIA ConnectX-7 family represents the next generation of RDMA network adapters, supporting both InfiniBand and Ethernet protocols at speeds up to 400Gb/s per port. The MCX755106AS-HEAT is a dual-port adapter delivering NDR (Next Data Rate) InfiniBand and 400GbE connectivity on both ports, built on PCIe Gen 5.0 technology for maximum host bandwidth.
ConnectX-7 introduces a comprehensive set of software-defined, hardware-accelerated networking, storage, and security capabilities. The dual-port configuration enables high-density compute nodes, redundant uplinks, or direct connections to multiple switches. Key innovations include inline IPsec/TLS/MACsec encryption offloads, advanced precision timing (IEEE 1588v2 Class C with 12ns accuracy), NVIDIA In-Network Computing engines, and NVMe-oF acceleration with GPUDirect support.
Key Features
- Dual-port 400Gb/s connectivity – 800Gb/s aggregate; NDR InfiniBand (IBTA Spec 1.5 compliant) and 400GbE support per port
- PCIe Gen 5.0 x16 host interface – Up to 128GB/s bidirectional bandwidth; backward compatible with PCIe Gen 4.0/3.0
- Hardware security offloads – Inline IPsec AES-GCM 128/256-bit, inline TLS AES-GCM 128/256-bit, inline MACsec AES-GCM 128/256-bit
- Platform security – Secure boot with hardware root-of-trust, secure firmware update, flash encryption, device attestation
- Advanced timing and synchronization – IEEE 1588v2 (any profile), G.8273.2 Class C, 12 nanosecond accuracy, line-rate hardware timestamp, SyncE, PTP-based packet pacing
- In-Network Computing acceleration – Collective operations offloads, NVIDIA SHARP support, rendezvous protocol offload, on-board memory
- GPUDirect RDMA and GPUDirect Storage – Direct GPU memory access for accelerated AI/ML workloads across both ports
- NVMe-oF offloads – NVMe over Fabrics target and initiator offloads, NVMe/TCP support
- RDMA with reliable transport – Extended Reliable Connected (XRC), Dynamically Connected Transport (DCT), on-demand paging (ODP)
- ASAP2 SDN acceleration – Open vSwitch offload, overlay network offload (VXLAN, GENEVE, NVGRE), programmable flexible parser
- Virtualization – SR-IOV, VirtIO acceleration, connection tracking (L4 firewall), flow mirroring, hierarchical QoS
- NVIDIA Multi-Host support – Connect up to 4 hosts to a single adapter (dependent on configuration)
- Redundancy and load balancing – Dual-port design enables link aggregation, failover, and multi-switch connectivity
Technology: In-Network Computing & Security Acceleration
ConnectX-7 elevates network intelligence with advanced offload engines that transform the network into a computational and security resource.
- In-Network Computing: Collective operations offload (MPI_Allreduce, etc.), NVIDIA SHARP, rendezvous offload, burst buffer offload
- Inline IPsec: AES-GCM 128/256-bit, IPsec over RoCE, line-rate encryption/decryption
- Inline TLS: AES-GCM 128/256-bit for secure application layer communication
- Inline MACsec: AES-GCM 128/256-bit for link-layer security
- Precision Timing: IEEE 1588v2 Class C (12ns accuracy), SyncE (G.8262.1), configurable PPS, time-triggered scheduling
- GPUDirect: RDMA and Storage – direct GPU memory access, bypassing CPU
Typical Deployments
- AI Factories and Large Language Model Training: Massive-scale distributed training with dual-port connectivity for north-south and east-west traffic, GPUDirect RDMA, and SHARP collective offloads.
- High Performance Computing (HPC): Scientific simulations, weather modeling, genomics requiring dual-port for high availability or multi-fabric connectivity.
- Secure Multi-Tenant Cloud Data Centers: Hardware-accelerated IPsec/TLS for zero-trust security with no CPU performance penalty, dual-port for redundant uplinks.
- NVMe-oF Storage Systems: High-performance shared storage with hardware offloads for NVMe/TCP and NVMe/RoCE, dual-port for storage controller redundancy.
- Telecom and Edge Computing: Precision timing (Class C) for 5G fronthaul and timing-sensitive infrastructure with redundant timing sources.
- Software-Defined Networking (SDN): ASAP2 acceleration for Open vSwitch with line-rate performance across both ports.
- Hyperscale Data Centers: Dual-port adapters for spine-leaf architectures, reducing number of adapters per server while maintaining redundancy.
Compatibility & Interoperability
- InfiniBand: Fully compatible with NVIDIA Quantum-2, Quantum-3 (when available), and legacy InfiniBand switches (NDR, HDR, EDR, FDR).
- Ethernet: Interoperable with 400GbE, 200GbE, 100GbE, 50GbE, 25GbE, 10GbE switches.
- PCIe: PCIe Gen 5.0 x16 – requires Gen 5.0 slot for full 400Gb/s per port performance; backward compatible with Gen 4.0 and Gen 3.0.
- CPU architectures: x86 (Intel/AMD), Arm, Power, GPU-accelerated platforms.
- Cabling: Passive copper DAC, active copper cables (ACC), active optical cables (AOC), optical transceivers – all with ESD protection.
Technical Specifications
| Parameter | Specification (MCX755106AS-HEAT) |
|---|---|
| Product Model | MCX755106AS-HEAT |
| Product Family | NVIDIA ConnectX-7 |
| Maximum Data Rate per Port | 400Gb/s (NDR InfiniBand or 400GbE) |
| Total Aggregate Bandwidth | 800Gb/s (2 x 400Gb/s) |
| Port Configuration | 2 ports (QSFP type - please confirm with ordering guide) |
| Host Interface | PCIe Gen 5.0 x16 (up to 128GB/s); compatible with Gen 4.0, 3.0 |
| InfiniBand Compliance | IBTA Specification 1.5, NDR/HDR/EDR/FDR |
| Ethernet Speeds Supported | 400/200/100/50/40/25/10/1 GbE |
| RDMA Support | Yes, with RoCE support |
| Transport Protocols | RC, UC, UD, XRC, DCT, Raw Datagram |
| Security Offloads | Inline IPsec (AES-GCM 128/256), inline TLS (AES-GCM 128/256), inline MACsec (AES-GCM 128/256) |
| Platform Security | Secure boot (hardware root-of-trust), secure firmware update, flash encryption, device attestation |
| Precision Timing | IEEE 1588v2 (any profile), G.8273.2 Class C, 12ns accuracy, SyncE (G.8262.1), PTP packet pacing |
| NVMe-oF Offloads | NVMe over Fabrics target/initiator, NVMe/TCP |
| Storage Protocols | SRP, iSER, NFS over RDMA, SMB Direct, T10-DIF signature handover |
| Virtualization | SR-IOV, VirtIO acceleration, connection tracking, flow mirroring, hierarchical QoS |
| Overlay Offloads | VXLAN, GENEVE, NVGRE encapsulation/decapsulation |
| ASAP2 SDN Acceleration | Open vSwitch offload, programmable flexible parser, header rewrite, hierarchical QoS |
| GPUDirect | GPUDirect RDMA, GPUDirect Storage |
| Multi-Host Support | NVIDIA Multi-Host (configuration dependent) |
| Remote Boot | InfiniBand, iSCSI, UEFI, PXE |
| Management | NC-SI, MCTP over SMBus/PCIe, PLDM (Monitor/Control/Firmware/Redfish/FRU), SPDM, SPI to flash, JTAG |
| Operating System Support | Linux (RHEL, Ubuntu), Windows, VMware ESXi, Kubernetes (in-box drivers for major OS) |
| Form Factor | Please confirm ordering information – HHHL, FHHL variants available |
| Standards Compliance | RoHS, IEEE 802.3bj/ba/by/ap, IEEE 1588v2, G.8273.2, G.8262.1 |
Note: Complete ordering information including specific form factor and connector type is available in the official NVIDIA ConnectX-7 user manuals. Please confirm exact specifications before ordering.
Why Choose ConnectX-7 MCX755106AS-HEAT
- Unmatched 400Gb/s per port throughput: Dual-port design delivers 800Gb/s aggregate for the most demanding AI and HPC workloads.
- PCIe Gen 5.0 readiness: Future-proof interface with 128GB/s host bandwidth – sufficient for dual 400Gb/s ports (requires efficient PCIe utilization).
- Zero-trust security without compromise: Inline IPsec/TLS/MACsec offloads encrypt all traffic at line rate on both ports with no CPU overhead.
- Industry-leading precision timing: 12ns accuracy, Class C compliant – ideal for 5G, finance, and distributed databases requiring redundant timing.
- Complete storage offload: NVMe-oF target and initiator acceleration reduces CPU load and improves storage latency across both ports.
- Redundancy and high availability: Dual-port design enables link aggregation, failover, and multi-switch connectivity for mission-critical deployments.
- Proven software ecosystem: Full driver support for Linux, Windows, VMware, Kubernetes, and all major HPC middleware.
Frequently Asked Questions
Q: What is the difference between MCX755106AS-HEAT and single-port ConnectX-7 adapters?
A: MCX755106AS-HEAT offers two independent 400Gb/s ports (800Gb/s aggregate), enabling redundant connections to two different switches or direct dual-fabric connectivity. Single-port versions like MCX755105AS-HEAT provide one 400Gb/s port for simpler topologies.
Q: Can both ports operate at full 400Gb/s simultaneously?
A: Yes, both ports can achieve line-rate 400Gb/s simultaneously. However, the PCIe Gen 5.0 x16 interface provides 128GB/s (approximately 1024Gb/s) bidirectional bandwidth, which supports the 800Gb/s aggregate transmit or receive with sufficient headroom.
Q: Does MCX755106AS-HEAT support both InfiniBand and Ethernet on each port independently?
A: Yes, ConnectX-7 adapters are VPI capable. Each port can be configured independently for InfiniBand or Ethernet, allowing mixed protocol deployments on the same adapter.
Q: What PCIe slot is required for full dual-port 400Gb/s performance?
A: A PCIe Gen 5.0 x16 slot is required for optimal performance. The adapter is backward compatible with PCIe Gen 4.0 x16, but aggregate throughput may be limited to approximately 400-500Gb/s due to host bandwidth constraints.
Q: What is the security advantage of inline offloads?
A: Inline IPsec/TLS/MACsec offloads encrypt and decrypt traffic directly on the adapter without CPU involvement, enabling full 400Gb/s secure networking on both ports with zero CPU penalty.
Q: What is the timing accuracy of IEEE 1588v2 on ConnectX-7?
A: ConnectX-7 achieves 12 nanosecond accuracy with G.8273.2 Class C compliance, suitable for the most demanding telecom and timing-sensitive applications across both ports.
Q: Does this adapter support GPUDirect across both ports?
A: Yes, ConnectX-7 fully supports GPUDirect RDMA and GPUDirect Storage on both ports, enabling direct GPU-to-GPU communication across multiple fabrics.
Installation & Handling Precautions
- ESD protection: Use grounded wrist straps and ESD-safe workstations when handling the adapter.
- PCIe Gen 5.0 signal integrity: Ensure server motherboard and slot meet Gen 5.0 signal quality requirements for optimal dual-port performance.
- Thermal management: Ensure adequate chassis airflow – dual-port ConnectX-7 adapters generate significant heat under full 800Gb/s aggregate load.
- Power requirements: PCIe slot must provide adequate power per PCIe CEM specifications for high-power operation. Verify server power delivery capability.
- Firmware updates: Keep adapter firmware current using NVIDIA Firmware Tools (MFT) for security patches and performance improvements.
Compatibility Matrix
| Platform / Switch Type | Compatibility | Notes |
|---|---|---|
| NVIDIA Quantum-2 / Quantum-3 InfiniBand switches | Full | NDR 400Gb/s, SHARP support on both ports |
| NVIDIA Spectrum-4 Ethernet switches | Full | 400GbE, RoCE, PTP timing on both ports |
| NVIDIA DGX H100 / DGX B200 systems | Full | GPUDirect RDMA and Storage across both ports |
| Standard PCIe Gen 5.0 servers (Dell, HPE, Supermicro, Lenovo, Cisco) | Full | Requires Gen 5.0 x16 slot for full speed |
| PCIe Gen 4.0 servers | Limited | Operates at Gen 4.0 speeds – sufficient for many single-port workloads |
Buyer Checklist
- Required per-port speed: 400Gb/s NDR – confirm network infrastructure supports NDR or 400GbE on both ports.
- Port count: Dual-port (MCX755106AS-HEAT) vs. single-port – determine redundancy or multi-fabric requirements.
- Host interface: Verify server has available PCIe Gen 5.0 x16 slot for maximum aggregate throughput.
- Protocol per port: InfiniBand, Ethernet, or mixed – adapter supports independent configuration.
- Security requirements: IPsec, TLS, or MACsec offload needs – confirm target use case across both ports.
- Timing requirements: PTP Class C (12ns) for telecom or financial applications – verify port-specific timing needs.
- Form factor: Confirm if HHHL or FHHL fits server chassis.
- Power and cooling: Verify server can provide sufficient power and cooling for dual-port operation.
- Cabling: Passive DAC (short reaches), active copper cables, AOC, or optical transceivers per port.
Hong Kong Starsurge Group Co., Limited
Founded in 2008, Starsurge is a technology-driven provider of network hardware, IT services, and system integration solutions. We serve global industries including government, healthcare, finance, manufacturing, AI, and education. Backed by an experienced sales and technical team, Starsurge supplies original NVIDIA networking products including ConnectX smart adapters, InfiniBand switches, cables, and transceivers. Our customer-first approach ensures reliable quality, responsive service, and efficient network infrastructure deployment worldwide.
이 제품에 대한 자세한 내용을 알고 싶습니다