NVIDIA Mellanox® MSB7890-ES2F — Switch-IB 2 EDR 36-Port Externally Managed InfiniBand Switch (1U, Dual PSU, P2C)

Technical Identity
The NVIDIA Mellanox® MSB7890-ES2F is a 1U externally managed InfiniBand switch built on the Switch-IB 2 ASIC. It provides 36 non-blocking EDR ports at 100 Gb/s per port over QSFP28 cages, delivering 7.2 Tb/s of aggregate bidirectional bandwidth with sub-100 ns port-to-port latency. The "E" suffix denotes Externally Managed (no on-switch CPU / no embedded subnet manager) — a managed subnet manager elsewhere in the fabric (UFM, OpenSM on a host, or a managed Switch-IB 2 peer) is required. The "S2F" suffix denotes Standard chassis, 2 hot-swap AC PSUs, and P2C airflow (Power-side intake → Connector-side exhaust, i.e. rear-to-front). Condition: New, factory-sealed.

Specifications

Brand / OEM: NVIDIA Mellanox
Part Number (MPN): MSB7890-ES2F
Product family: Switch-IB 2 (SB7800 series)
Switch ASIC: Switch-IB 2
Port count: 36 ports
Port speed: EDR InfiniBand — 100 Gb/s per port (4× 25 Gb/s NRZ lanes)
Aggregate throughput: 7.2 Tb/s bidirectional
Latency: sub-100 ns port-to-port (cut-through)
Connector type: QSFP28
Protocol: InfiniBand only (NOT VPI, NOT Ethernet)
Management: Externally Managed — no embedded CPU, no internal subnet manager
Form factor: 1U standard 19" rack-mount
Power supplies: 2× hot-swap AC PSUs (1+1 redundant)
Airflow: P2C — Power-side intake to Connector-side exhaust (rear-to-front)
Cooling: Hot-swap redundant fan modules matched to airflow direction
Routing features: Adaptive Routing, Congestion Control, SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) acceleration
Condition: New, factory-sealed


What This Is NOT

NOT a managed switch with internal subnet manager. The "E" suffix means Externally Managed — a separate subnet manager (UFM appliance, OpenSM host, or managed peer switch) is mandatory to bring the fabric up.
NOT an Ethernet switch. Switch-IB 2 is InfiniBand-only silicon — no Ethernet port mode, no VPI, no MLAG.
NOT HDR or NDR. EDR class (100 Gb/s per port) — do not confuse with Quantum (HDR 200 Gb/s) or Quantum-2 (NDR 400 Gb/s) families.
NOT QSFP / QSFP+ / QSFP56 / OSFP. Connector is QSFP28 — use QSFP28 EDR cables and transceivers only.
NOT bundled with cables, transceivers, rail kit, or console cable unless explicitly listed in the offer.


Compatibility & Constraints

Subnet manager: Mandatory external SM. Common deployments use a managed Switch-IB 2 / Quantum peer running embedded SM, NVIDIA UFM Enterprise / UFM Cyber-AI, or OpenSM on a Linux host attached to the fabric.
Compatible host adapters: NVIDIA Mellanox ConnectX-4 EDR (MCX455A-ECAT, MCX456A-ECAT), ConnectX-5 EDR / VPI in IB mode (MCX555A-ECAT, MCX556A-ECAT), ConnectX-6 VPI in IB mode at EDR rate (MCX653105A/106A in EDR mode), and BlueField-1 / BlueField-2 EDR DPUs. Higher-class adapters (HDR/NDR) auto-negotiate down to EDR.
Cabling: QSFP28 EDR DACs (e.g. MCP1600-Exxx series), QSFP28 EDR AOCs (e.g. MFA1A00-Exxx series), and QSFP28 EDR optical modules (MMA1B00-E100, MMS1C10-CM, etc.). Use 100GbE QSFP28 cables only when EDR-rated.
Topology: Suited for spine/leaf, fat-tree, dragonfly, and torus fabrics. Adaptive Routing and SHARP acceleration significantly improve all-reduce performance for distributed training and HPC collectives.
Power input: 100–240 VAC, 50/60 Hz — verify rack PDU phase / IEC C13 / C14 inlet availability before installation.
Airflow direction: P2C must match the cold-aisle / hot-aisle layout of the rack. Mixing P2C and C2P units in the same rack column causes recirculation — plan airflow before deployment.
Firmware: Switch-IB 2 firmware update is performed via the fabric (in-band) from a managed peer or UFM — not via on-switch console (no CPU / no Onyx CLI on Externally Managed SKUs).


Frequently Asked Questions (FAQ)
Q. What is the NVIDIA Mellanox MSB7890-ES2F?
The MSB7890-ES2F is a 1U Switch-IB 2 InfiniBand switch with 36 EDR ports (100 Gb/s) on QSFP28 cages. The "E" suffix means it is Externally Managed — it has no on-switch CPU and requires an external subnet manager. The "S2F" suffix means Standard chassis, dual hot-swap AC PSUs, and P2C (rear-to-front) airflow.
Q. Does the MSB7890-ES2F run Ethernet or VPI?
No. Switch-IB 2 silicon is InfiniBand only. There is no Ethernet port mode and no VPI dual-protocol capability. For Ethernet at 100 GbE QSFP28, use the Spectrum family (e.g. SN2700 / SN2410).
Q. Do I need a separate subnet manager?
Yes — mandatory. The "E" suffix means there is no embedded subnet manager. Deploy one of: NVIDIA UFM Enterprise / UFM Cyber-AI appliance, OpenSM on a Linux host attached to the fabric, or a managed Switch-IB 2 / Quantum peer running its embedded SM.
Q. Which cables and adapters work with this switch?
Use QSFP28 EDR DACs, AOCs, or transceivers (Mellanox MCP1600-Exxx, MFA1A00-Exxx, MMA1B00-E100, etc.). Compatible host adapters include ConnectX-4 EDR, ConnectX-5 EDR / VPI-in-IB-mode, ConnectX-6 VPI in EDR mode, and BlueField-1 / -2 DPUs. Higher-class HDR / NDR adapters auto-negotiate down to EDR.
Q. What does "P2C airflow" mean for installation?
P2C = Power-side intake to Connector-side exhaust. Air enters the rear (PSU side) and exhausts at the front (port side). This is the opposite direction of the standard "front-to-back" hot-aisle convention used by many servers — verify your rack airflow plan before installation to avoid hot-air recirculation.
Q. Is this unit new or refurbished?
This SKU is supplied new, factory-sealed in original NVIDIA Mellanox packaging where included.

Why buy from T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS is a Europe-based NVIDIA Mellanox specialist serving HPC, AI, and enterprise data-center procurement teams since 2007. Every switch is identified by MPN, NVIDIA P/N, ASIC generation, airflow direction, and management mode before it ships, so the unit you receive matches your fabric design — not just the catalog photo. New units arrive factory-sealed; refurbished units are bench-tested for link establishment, PSU redundancy, and fan/airflow direction. EU stock with worldwide express shipping, multi-currency B2B invoicing, and direct technical contact with our engineering team for subnet-manager and fabric-topology planning at no extra cost.