NVIDIA Mellanox® 90Y3453 — IBM Flex System IB6131 InfiniBand Switch Module
Technical Identity
The NVIDIA Mellanox® 90Y3453 is the IBM FRU part number for the IBM Flex System IB6131 InfiniBand Switch, a chassis-resident I/O module built on Mellanox SwitchX-2 silicon for the IBM Flex System Enterprise Chassis (8721 / 7893). It is a true InfiniBand fabric switch — not Ethernet, not converged — delivering up to FDR 56 Gb/s per port and backward-compatible with FDR10 (40 Gb/s), QDR (40 Gb/s), and DDR (20 Gb/s) endpoints. The module presents 14 internal ports to the chassis midplane (one per compute-node bay, paired with an IB mezzanine adapter such as the IB6132 2-port FDR card) and 18 external QSFP+ uplink ports for spine, gateway, or director-class aggregation. An embedded subnet manager makes the IB6131 a fully self-contained fabric element — no external SM host required for small to mid-size clusters.
Specifications
- OEM / Brand: NVIDIA Mellanox (IBM-branded FRU)
- IBM FRU PN (MPN): 90Y3453
- IBM option PN: 90Y3450
- Product family: IBM Flex System IB6131 InfiniBand Switch
- Silicon: Mellanox SwitchX-2
- Protocol: InfiniBand (IB-only — no Ethernet/RoCE)
- Top line rate: FDR 56 Gb/s per port
- Backward compat: FDR10 40 Gb/s, QDR 40 Gb/s, DDR 20 Gb/s, SDR 10 Gb/s
- Internal ports: 14 × FDR-capable (chassis midplane, one per node bay)
- External ports: 18 × QSFP+ uplinks (FDR14)
- Switching capacity: 1.8 Tb/s non-blocking, cut-through
- Subnet manager: embedded SM (up to 648 nodes)
- Management: CLI (SSH/Telnet), Web UI, SNMP, IPMI passthrough via CMM
- Form factor: IBM Flex System I/O bay module (hot-swap) — NOT a 1U rack switch
- Latency: < 200 ns port-to-port (FDR)
- Condition: New, factory-sealed
What This Is NOT
- NOT a standalone 1U rack switch. The IB6131 is a chassis I/O module for the IBM Flex System Enterprise Chassis (8721 / 7893). It will not function outside that chassis.
- NOT an Ethernet or VPI switch. InfiniBand only. For Ethernet uplinks in Flex System, use the EN4093 / SI4093 modules.
- NOT an HDR or EDR-generation switch. Top rate is FDR 56 Gb/s; for EDR 100 Gb/s use Switch-IB (SB7700), for HDR 200 Gb/s use Quantum (QM8700), for NDR 400 Gb/s use Quantum-2 (QM9700).
- NOT bundled with QSFP+ DACs, AOCs, or transceivers. External uplinks need separately ordered FDR-rated QSFP+ cables (MC2207-series).
- NOT bundled with IB mezzanine adapters. Each compute node needs its own IB6132 mezzanine card.
Compatibility
- Host chassis: IBM Flex System Enterprise Chassis (8721-HC1, 7893-92X, 7893-A2X). One IB6131 occupies a single I/O bay.
- Compute node mezzanine: IBM Flex System IB6132 2-port FDR Adapter (PN 90Y3454 / FRU 90Y3486) — Mellanox ConnectX-3 based.
- Supported compute nodes: x220, x240, x440, p24L, p260, p270, p460 (with IB6132 populated).
- External cabling: Mellanox MC2207-series passive copper QSFP+ DACs, MC2206-series active fiber for FDR; FDR10/QDR cables for back-compat speeds.
- Spine / aggregation switches it pairs with: Mellanox SX6036 / SX6025 (FDR managed) and SX6512 / SX6536 director chassis. Also links to EDR Switch-IB (SB7700/7800) at FDR speeds via FDR-rated cables.
- Subnet manager: embedded SM on the IB6131, OpenSM on a Linux host, or NVIDIA UFM for large fabrics.
- Cluster middleware: MLNX OFED, RDMA verbs, OpenMPI / Intel MPI / IBM Platform MPI, NFS-RDMA, GPFS / Spectrum Scale RDMA, Lustre with o2ib.
Deployment Scenarios
- HPC inside Flex System: dual-rail IB fabric across a chassis of x240 / p260 nodes for MPI workloads.
- GPFS / Spectrum Scale storage fabric: RDMA back-end between Flex compute and external storage controllers.
- Database scale-out: low-latency interconnect for Oracle RAC, DB2 pureScale, SAP HANA scale-out on Flex nodes.
- Lab / staging clusters: self-contained chassis fabric with embedded SM, no external IB switch required.
FAQ
Q: What exactly is the 90Y3453?
A: 90Y3453 is the IBM FRU part number for the IBM Flex System IB6131 InfiniBand Switch — a Mellanox SwitchX-2-based FDR 56Gb/s InfiniBand I/O module for the IBM Flex System Enterprise Chassis. The IBM option PN is 90Y3450.
Q: Is this an Ethernet switch?
A: No. The IB6131 is InfiniBand only. It does not run Ethernet, RoCE, or VPI dual-protocol mode. For Ethernet uplinks in Flex System, use the EN4093R or SI4093 modules.
Q: What speed does it support?
A: Top line rate is FDR 56 Gb/s per port, backward-compatible with FDR10 (40 Gb/s), QDR (40 Gb/s), DDR (20 Gb/s), and SDR (10 Gb/s) — useful when mixing older ConnectX-2 / ConnectX-3 nodes.
Q: How many ports does it have?
A: 14 internal FDR ports (one per compute-node bay through the chassis midplane) plus 18 external QSFP+ uplink ports.
Q: Can I use the 90Y3453 outside an IBM Flex System chassis?
A: No. It is a chassis I/O module dependent on the IBM Flex System Enterprise Chassis (8721 / 7893). It is not a standalone rack switch.
Q: Does it include a subnet manager?
A: Yes. The IB6131 has an embedded subnet manager capable of running fabrics up to ~648 nodes. For larger fabrics, run OpenSM on a host or deploy NVIDIA UFM.
Q: What IB adapter card pairs with this on compute nodes?
A: The IBM Flex System IB6132 2-port FDR Adapter (PN 90Y3454 / FRU 90Y3486), a Mellanox ConnectX-3 mezzanine card.
Q: Will it interoperate with newer EDR / HDR switches?
A: Yes, at FDR speed. Links from IB6131 QSFP+ uplinks to EDR Switch-IB (SB7700) or HDR Quantum (QM8700) auto-negotiate to FDR 56 Gb/s when FDR-rated cables are used.
Q: Is this a new or used unit?
A: New, factory-sealed, bench-tested by T.E.S IT-SOLUTIONS before shipment.
Why T.E.S IT-SOLUTIONS
Every IB6131 module ships from EU stock after bench testing by T.E.S IT-SOLUTIONS: power-on validation in a live Flex System chassis, port-by-port link verification on internal and external ports, firmware check, and cosmetic inspection. Anti-static packaging, global B2B shipping, and direct technical consultation for IBM Flex System / PureFlex InfiniBand fabric design, mixed-speed coexistence with EDR/HDR spines, and subnet-manager planning. Ideal for customers maintaining or expanding existing Flex System HPC, GPFS, or database clusters where like-for-like FDR replacement is required. See tes-itsolutions.com.
