Mellanox SB7890-36 - 36 x 100GbE EDR InfiniBand Switch, 1U, Externally Managed (REF)

Mellanox SB7890-36 is a high-performance enterprise InfiniBand switch with 36 x 100GbE QSFP28 EDR ports. It delivers 7.2 Tb/s non-blocking bandwidth and ultra-low 90 ns latency, optimized for HPC, AI/ML, and large-scale data centre environments.

With redundant hot-swappable power supplies and fans, a compact 1U form factor, and externally managed capabilities, it provides reliable, high-density connectivity. Advanced features such as SHARP™ in-network computing, adaptive routing, congestion control, and VL2VL mapping allow administrators to optimize network performance, reduce CPU overhead, and maintain scalability for demanding workloads.

KEY FEATURES:

• Access Ports: 36 x 100GbE QSFP28 InfiniBand EDR
• Switching Capacity: 7.2 Tb/s non-blocking
• Latency: 90 ns port-to-port
• Uplink Options: All ports are access ports; supports breakout cables for 4x25GbE per port
• System Memory: 16 GB DDR4
• Flash Storage: 32 GB
• CPU: 2 x 16-core ARM64 processors
• Forwarding Rate: Up to 7.2 Bpps
• Packet Buffer: 16 MB shared
• MAC Address Table: 192K entries
• VLAN Support: Supports InfiniBand partitions (P_Key)
• Stacking / Fabric: Supports fabric topologies like Fat Tree, Dragonfly+, SlimFly
• Resiliency: Redundant, hot-swappable AC power supplies and fans
• Airflow: Connector side inlet (C2P) or outlet (P2C)
• Form Factor: 1U rack-mountable
• Condition: Refurbished
  

TECHNICAL SPECIFICATIONS:

HARDWARE:

PERFORMANCE:

SECURITY:

MANAGEMENT:

PHYSICAL & ENVIRONMENTAL:

ADVANCED FEATURES:

SHARP™ In-Network Computing:

SHARP™ accelerates collective communications such as MPI and SHMEM directly in the network hardware. By offloading operations from the CPU, it significantly reduces processing overhead and latency, improving overall application performance. This is especially beneficial for HPC and AI/ML workloads that require synchronized operations across large clusters. With SHARP™, data movement is optimized, enabling faster computation cycles and efficient resource utilization, which reduces energy consumption and improves system throughput across the fabric.

Adaptive Routing & Congestion Control:

Adaptive routing dynamically spreads traffic across multiple paths, automatically avoiding congestion hotspots. Coupled with in-network congestion control, this ensures high throughput and predictable performance even in complex, large-scale network topologies such as Fat Tree, Dragonfly+, and SlimFly. The system continuously monitors link utilization and fabric conditions to redirect traffic in real-time, maintaining minimal latency and preventing bottlenecks, which is critical for performance-sensitive HPC, AI/ML, and large-scale storage environments.

VL2VL Mapping:

VL2VL mapping provides flexible and efficient virtual lane utilization across the network. By dynamically mapping virtual lanes to physical lanes, the feature minimizes head-of-line blocking, improves traffic isolation, and optimizes throughput. It is particularly valuable for multi-tenant, multi-job, or mixed workload environments where efficient lane allocation ensures fair bandwidth distribution. VL2VL ensures that applications and jobs experience consistent, predictable performance, reducing contention and maximizing fabric efficiency across diverse network architectures.

Telemetry and Monitoring:

The SB7890-36 supports advanced telemetry for comprehensive real-time monitoring of the entire fabric. Administrators can track throughput, latency, congestion, and hardware health using performance counters, SNMP, and remote diagnostics. Data from the switch is continuously analyzed to detect anomalies and optimize traffic flows, enabling proactive management. Telemetry helps reduce downtime by predicting potential issues, supporting automated alerts, and providing insights that improve decision-making for HPC clusters, AI/ML workloads, and data centre operations.

Energy-Efficient Design:

The SB7890-36 is designed to operate efficiently at only 122 W typical under full load. Intelligent fan control adjusts airflow dynamically based on thermal conditions, reducing power consumption and cooling costs. Combined with hardware-level optimizations, this energy-efficient design helps lower operational costs in data centres and HPC clusters. Despite its low power footprint, the switch maintains full performance and high throughput, ensuring that energy savings do not compromise reliability, performance, or scalability in demanding workloads.

Cluster and HPC Optimization:

Topology-aware routing and in-network offloads enhance performance for HPC clusters and AI/ML environments. Congestion control, adaptive routing, and SHARP™ support provide ultra-low latency and high throughput across large-scale clusters. This optimization ensures that computational tasks are completed faster and with minimal communication delays. The switch efficiently handles large numbers of nodes and multi-job environments, maintaining predictable performance while scaling seamlessly, making it ideal for research, simulation, analytics, and AI/ML training workloads.

Redundancy and Reliability:

The switch includes dual hot-swappable AC power supplies and N+1 fans for uninterrupted operation. Redundant hardware components and externally managed capabilities allow proactive failover management, ensuring no single point of failure affects fabric performance. The system continuously monitors power and thermal conditions, triggering alerts and adjustments as needed. High availability ensures sustained connectivity for critical workloads in HPC, AI/ML, and large-scale data centres, minimizing downtime and maintaining consistent throughput under demanding operational conditions.

Fabric Automation:

SB7890-36 supports comprehensive automation for provisioning, monitoring, and fabric management. Administrators can automate switch configuration, health checks, and event-triggered alerts to minimize manual intervention. Automation simplifies deployment of large fabrics, reduces operational errors, and improves consistency across clusters. By integrating with orchestration tools and monitoring platforms, it provides real-time updates on performance and utilization. This enables faster troubleshooting, proactive maintenance, and efficient allocation of network resources across large-scale HPC and AI/ML environments.

Enhanced Security Integration:

Security is integrated at both firmware and fabric levels. Features include secure boot, role-based access control, encryption, and audit logging. Administrators can enforce policies centrally, monitor unauthorized access, and integrate logs with security dashboards for compliance. These measures protect sensitive HPC, AI/ML, and research workloads from tampering or eavesdropping. With robust authentication and encrypted management protocols, the SB7890-36 ensures the network fabric remains secure, reliable, and resilient against threats while maintaining high performance and availability.

ORDERING INFORMATION:

SUPPORT:

All refurbished products sold on Grabnpay are tested, verified, and delivered in good working condition. Our team provides assistance with installation, configuration, and troubleshooting to ensure smooth operations. Please note that for refurbished products, no manufacturer warranty, vendor support, or license coverage applies. For any issues, our technical support team will guide you with solutions and recommend suitable replacements or upgrades if required.