Mellanox MQM8790-HS2F

Detailed Feature Explanation

1. ConnectX-6 Dx ASIC: This is the engine of the card, providing:

   • High Throughput: Capable of handling 200Gb/s per port, saturating the link with minimal CPU usage.

   • Low Latency: Extremely fast packet processing, crucial for financial trading, AI, and database applications.

2. RDMA over Converged Ethernet (RoCE): This is the card's standout feature.

   • RDMA (Remote Direct Memory Access) allows one computer to directly access the memory of another without involving the CPU of either machine.

   • This drastically reduces latency and CPU overhead, which is essential for:

     • AI/ML Clusters: Accelerating communication between GPU servers (e.g., with NVIDIA's NCCL).

     • Distributed Storage & HCI: Providing fast access to network-attached storage (like VAST Data, Dell ECS).

     • Databases: Improving performance for distributed databases.

3. CPU Offloads: The card has dedicated hardware to handle complex tasks, freeing up the server's CPU cores for applications.

   • NVMe-oF Target Offload: Allows the server to act as a high-performance storage array.

   • IPsec In-line Encryption: Offloads encryption/decryption processes, providing high-speed data security without taxing the CPU.

   • VXLAN Offloading: Handles network overlay encapsulation/de-capsulation in hardware.

4. SR-IOV (Single Root I/O Virtualization): Allows the physical NIC to be partitioned into multiple virtual NICs (vNICs) that can be assigned directly to virtual machines (VMs). This provides near-native network performance for VMs.

Common Use Cases

• AI & Machine Learning: Installed in training servers to enable high-speed RoCE communication for multi-node GPU training.

• High-Performance Computing (HPC): Connects compute nodes in a cluster for fast message passing (e.g., with MPI).

• Hyper-Converged & Software-Defined Storage: Provides the fast network backbone for storage traffic (e.g., Ceph, vSAN).

• Cloud and Telco Data Centers: Used for high-performance NFV (Network Functions Virtualization) and secure, encrypted multi-tenancy.

• High-Frequency Trading: Where the lowest possible latency is non-negotiable.

• Data Analytics: Accelerating platforms like Apache Spark.

How it Relates to the MSB7890-ES2F Switch

These two products are complementary and are often used together to build a complete high-performance fabric.

• The MQM8790-HS2F (NIC) goes inside the server.

• The MSB7890-ES2F (Switch) sits in the rack and connects all the servers.

A typical AI cluster setup:

1. Each AI server (with multiple GPUs) has an MQM8790-HS2F NIC.

2. Each NIC connects via a 200G DAC or AOC cable to the MSB7890-ES2F Switch.

3. The switch, running RoCE-enabled protocols, provides the low-latency, high-bandwidth fabric that allows all the servers to communicate and train a single AI model collaboratively.

   
   

   The Mellanox MQM8790-HS2F is a high-performance 200G Ethernet NIC based on the ConnectX-6 Dx ASIC. Its primary value is enabling RDMA (RoCE) and advanced CPU offloads, which are critical for achieving maximum performance in AI, HPC, storage, and cloud data centers. It is the endpoint that connects servers to the high-speed network fabric provided by switches like the MSB7890-ES2F.