 
|
NVIDIA ConnectX-6 MCX653105A-HDAT 200Gb/s InfiniBand & Ethernet Adapter
Λεπτομέρειες:
| Μάρκα: | Mellanox |
| Αριθμό μοντέλου: | MCX653105A-HDAT |
| Έγγραφο: | connectx-6-infiniband.pdf |
Πληρωμής & Αποστολής Όροι:
| Ποσότητα παραγγελίας min: | 1 ΤΕΜ |
|---|---|
| Τιμή: | Negotiate |
| Συσκευασία λεπτομέρειες: | Εξωτερικό κουτί |
| Χρόνος παράδοσης: | Βασισμένο στο αποθεματικό |
| Όροι πληρωμής: | T/T |
| Δυνατότητα προσφοράς: | Προμήθεια ανά έργο/παρτίδα |
|
Λεπτομερής ενημέρωση |
|||
| Κατάσταση προϊόντων: | Στοκ | Εφαρμογή: | Υπηρέτης |
|---|---|---|---|
| Κατάσταση: | ΝΕΟ ΚΑΙ ΠΡΩΤΟΤΥΠΟ | τύπος: | Ενσύρματο |
| Μέγιστη ταχύτητα: | Έως 200 GB/S | Σύνδεσμος Ethernet: | QSFP56 |
| Μοντέλο: | MCX653105A-HDAT | ||
| Επισημαίνω: | NVIDIA ConnectX-6 network adapter,200Gb/s InfiniBand Ethernet card,MCX653105A-HDAT network card with warranty |
||
Περιγραφή προϊόντων
NVIDIA ConnectX-6 MCX653105A-HDAT 200Gb/s InfiniBand & Ethernet Adapter
NVIDIA Original ConnectX-6 200Gb/s HDR InfiniBand + Ethernet
MPN: MCX653105A-HDAT
The NVIDIA ConnectX-6 MCX653105A-HDAT is a high-performance smart network adapter card delivering up to 200Gb/s throughput for both InfiniBand and Ethernet networks. Featuring a single QSFP56 port, PCIe Gen 4.0 x16 host interface, and NVIDIA's industry-leading In-Network Computing offloads, this adapter is purpose-built for HPC clusters, AI training platforms, hyperscale data centers, and storage infrastructures requiring ultra-low latency, high message rates, and exceptional CPU offload efficiency.
MPN: MCX653105A-HDAT
The NVIDIA ConnectX-6 MCX653105A-HDAT is a high-performance smart network adapter card delivering up to 200Gb/s throughput for both InfiniBand and Ethernet networks. Featuring a single QSFP56 port, PCIe Gen 4.0 x16 host interface, and NVIDIA's industry-leading In-Network Computing offloads, this adapter is purpose-built for HPC clusters, AI training platforms, hyperscale data centers, and storage infrastructures requiring ultra-low latency, high message rates, and exceptional CPU offload efficiency.
200Gb/s
Port Speed
InfiniBand & Ethernet
Protocol Support
215M
Messages/Sec
PCIe 4.0 x16
Host Interface
64GB/s
Bandwidth
RDMA
Zero CPU Copy
XTS-AES
Hardware Encryption
Product Overview
The NVIDIA ConnectX-6 family represents a breakthrough in smart network adapter design, integrating advanced offload engines that transform the network into a computational resource. The MCX653105A-HDAT is a single-port adapter card capable of 200Gb/s HDR (High Data Rate) InfiniBand and 200GbE Ethernet, providing exceptional performance for the most demanding workloads.
Leveraging NVIDIA In-Network Computing technology, ConnectX-6 offloads communication and computation tasks from the CPU to the adapter, reducing latency, increasing throughput, and freeing processor cores for application logic. Key capabilities include RDMA (Remote Direct Memory Access), NVMe-over-Fabrics target and initiator offloads, collective operations offload, block-level XTS-AES hardware encryption, and advanced virtualization features such as SR-IOV with up to 1K virtual functions.
Key Features
- Up to 200Gb/s connectivity per port – InfiniBand HDR and 200GbE support
- PCIe Gen 4.0 x16 host interface – Up to 64GB/s bidirectional bandwidth; backward compatible with PCIe Gen 3.0
- Ultra-low latency and up to 215 million messages per second
- In-Network Computing acceleration – Collective operations offload, vector collective operations, MPI tag matching, rendezvous protocol offload
- RDMA with end-to-end packet-level flow control – Compliant with IBTA specification
- NVMe-over-Fabrics offloads – Target and initiator offloads for optimized storage performance
- Block-level XTS-AES 256/512-bit hardware encryption – FIPS compliant options available
- Hardware-based I/O virtualization – SR-IOV with up to 1K virtual functions, 8 physical functions per host
- Overlay network offloads – VXLAN, NVGRE, Geneve encapsulation/decapsulation hardware offload
- NVIDIA GPUDirect RDMA – Direct GPU memory access for accelerated AI/ML workloads
- Dynamically Connected Transport (DCT) and Extended Reliable Connected (XRC) transport
- On demand paging (ODP) – Allows host memory paging for InfiniBand memory regions
- RoHS compliant and ODCC (Open Data Center Committee) compatible
Technology: In-Network Computing Acceleration
ConnectX-6 adapters are foundational to NVIDIA Quantum InfiniBand platform, delivering the world's only fully offloadable in-network computing acceleration technology. This offload engine enables the network fabric to actively participate in computation, dramatically improving performance for HPC, AI, and cloud infrastructures while reducing cost and complexity.
- Collective offloads: MPI_Allreduce, MPI_Barrier, MPI_Alltoall, etc.
- Vector collectives: Offloads for vector operations
- MPI tag matching: Hardware offload for message matching
- Rendezvous offload: Efficient large message handling
- Burst buffer offload: Accelerated I/O for burst buffers
- In-Network Memory: Registration-free RDMA memory access
Typical Deployments
- High Performance Computing (HPC): Large-scale clusters running MPI-based simulations, weather modeling, genomics, and computational fluid dynamics.
- Artificial Intelligence & Machine Learning: Distributed training of deep neural networks requiring high throughput and low latency between GPU nodes.
- NVMe-over-Fabrics Storage: High-performance storage arrays providing block access to NVMe devices over RDMA-capable networks.
- Hyperscale Cloud and Data Centers: Virtualized environments requiring advanced SR-IOV, overlay network offloads, and efficient I/O consolidation.
- GPU-Accelerated Computing: Servers with NVIDIA GPUs utilizing GPUDirect RDMA for direct GPU-to-GPU communication.
Compatibility & Interoperability
- InfiniBand: Fully compatible with NVIDIA Quantum, Quantum-2, and earlier InfiniBand switches (up to 200Gb/s HDR, HDR100, EDR, FDR, QDR).
- Ethernet: Interoperable with 200GbE, 100GbE, 50GbE, 40GbE, 25GbE, 10GbE, and 1GbE switches.
- Passive copper cables, active optical cables (AOC), and optical transceivers – with ESD protection.
- CPU architectures: x86 (Intel/AMD), Power, Arm, GPU, and FPGA-based platforms.
- PCIe compatibility: PCIe Gen 4.0 x16, also compatible with PCIe Gen 3.0, 2.0, 1.1.
Technical Specifications
| Parameter | Specification (MCX653105A-HDAT) |
|---|---|
| Product Model | MCX653105A-HDAT |
| Product Family | NVIDIA ConnectX-6 |
| Maximum Data Rate | 200Gb/s (InfiniBand HDR or 200GbE) |
| Port Configuration | 1x QSFP56 |
| Host Interface | PCIe Gen 4.0 x16 (up to 64GB/s); backward compatible with PCIe Gen 3.0, 2.0, 1.1 |
| Message Rate | Up to 215 million messages per second |
| Latency | Extremely low (sub-microsecond) |
| InfiniBand Compliance | IBTA Specification 1.3, HDR/HDR100/EDR/FDR/QDR |
| Ethernet Speeds Supported | 200/100/50/40/25/10/1 GbE |
| RDMA Support | Yes, with end-to-end packet-level flow control |
| NVMe-oF Offloads | Target and initiator offloads |
| Encryption | Block-level XTS-AES 256/512-bit hardware encryption |
| Virtualization | SR-IOV (up to 1K VFs), 8 physical functions per host, NPAR |
| Overlay Offloads | VXLAN, NVGRE, Geneve encapsulation/decapsulation |
| GPUDirect | NVIDIA GPUDirect RDMA (PeerDirect) |
| Transport Protocols | DCT, XRC, RC, UC, UD, Raw Datagram |
| Remote Boot | InfiniBand, Ethernet, iSCSI, UEFI, PXE |
| Form Factor | Low-profile stand-up PCIe (tall bracket mounted, short bracket included) |
| Dimensions (without bracket) | 167.65mm x 68.90mm |
| Operating System Support | RHEL, SLES, Ubuntu, other major Linux distros, Windows Server, FreeBSD, VMware, OpenFabrics (OFED/WinOF-2) |
| Management | NC-SI, MCTP over SMBus/PCIe, PLDM for Monitor/Control and Firmware Update |
| Standards Compliance | RoHS, ODCC, IEEE 802.3bj/ba/by/ae/ap, IEEE 1588v2, Jumbo frame (9.6KB) |
Selection Guide – ConnectX-6 Adapters
| Ordering Part Number | Ports | Speed | Host Interface | Form Factor | Description |
|---|---|---|---|---|---|
| MCX653105A-HDAT | 1x QSFP56 | 200Gb/s | PCIe Gen 4.0 x16 | Stand-up PCIe | Single-port HDR, InfiniBand & Ethernet |
| MCX653106A-HDAT | 2x QSFP56 | 200Gb/s | PCIe Gen 4.0 x16 | Stand-up PCIe | Dual-port HDR |
| MCX653105A-ECAT | 1x QSFP56 | 100Gb/s | PCIe Gen 4.0 x16 | Stand-up PCIe | Single-port HDR100 |
| MCX653435A-HDAT | 1x QSFP56 | 200Gb/s | PCIe Gen 4.0 x16 | OCP 3.0 SFF | OCP form factor variant |
Why Choose ConnectX-6 MCX653105A-HDAT
- Unmatched performance: 200Gb/s line rate with full hardware offloads ensures maximum throughput for data-intensive applications.
- CPU offload efficiency: Offload transport operations, collective communications, encryption, and storage protocols – saving CPU cycles for application logic.
- Future-ready PCIe Gen 4.0: Double the bandwidth of PCIe 3.0, ensuring no bottleneck between adapter and host memory.
- Secure by design: Hardware-based block encryption protects data in transit without performance penalty.
- Flexible deployment: Single port simplifies cabling for edge nodes or storage targets while delivering full 200Gb/s throughput.
- Broad software ecosystem: Full driver support for all major operating systems and HPC middleware (OpenMPI, MVAPICH, MPICH, etc.).
Frequently Asked Questions
Q: What is the difference between MCX653105A-HDAT and MCX653106A-HDAT?
A: MCX653105A-HDAT has a single QSFP56 port; MCX653106A-HDAT has two ports for dual 200Gb/s connectivity. Choose single-port for storage nodes or edge servers where one high-speed link suffices.
Q: Does this adapter support both InfiniBand and Ethernet?
A: Yes, ConnectX-6 adapters support both protocols. The adapter automatically configures based on the network environment. You can also use it for Ethernet with RoCE (RDMA over Converged Ethernet).
Q: Is the adapter compatible with PCIe Gen 3.0 servers?
A: Yes, the MCX653105A-HDAT is backward compatible with PCIe Gen 3.0, 2.0, and 1.1. When installed in a PCIe 3.0 slot, it will operate at Gen 3.0 speeds (8 GT/s).
Q: What is GPUDirect RDMA?
A: GPUDirect RDMA allows direct memory access between NVIDIA GPUs and the ConnectX-6 adapter, bypassing CPU memory. This dramatically reduces latency for GPU-to-GPU communication in AI/ML training clusters.
Q: Does the adapter support NVMe-over-Fabrics?
A: Yes, ConnectX-6 provides hardware offloads for both NVMe-oF initiator and target, improving storage performance and reducing CPU overhead.
Q: What operating systems are supported?
A: Major Linux distributions (RHEL, SLES, Ubuntu), Windows Server, FreeBSD, and VMware ESXi. NVIDIA provides OFED (OpenFabrics Enterprise Distribution) drivers.
Buyer Checklist
- Required port speed: 200Gb/s vs. 100Gb/s – confirm network infrastructure supports HDR/200GbE.
- Number of ports: single-port (MCX653105A-HDAT) or dual-port (MCX653106A-HDAT).
- Host interface: Verify server has available PCIe Gen 4.0 x16 slot for optimal performance.
- Protocol: InfiniBand or Ethernet – adapter supports both, but cabling and switch must match.
- Operating system: Confirm driver availability (Linux, Windows, VMware, FreeBSD).
- Cable type: Passive DAC (≤5m), active optical cable (longer reaches), or optical transceivers.
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 InfiniBand adapters, switches, cables, and transceivers. Our customer-first approach ensures reliable quality, responsive service, and efficient network infrastructure deployment worldwide.
Θέλετε να μάθετε περισσότερες λεπτομέρειες σχετικά με αυτό το προϊόν