NVIDIA Quantum switch is part of NVIDIA’s InfiniBand networking portfolio, designed for high-performance computing (HPC), artificial intelligence (AI), and hyperscale data center applications. Unlike the Spectrum series, which uses Ethernet, Quantum switches leverage InfiniBand, a high-speed, low-latency interconnect technology optimized for extreme performance and scalability. These switches are built to handle massive data throughput, making them ideal for workloads like AI training, scientific simulations, and cloud-native supercomputing.

Key Features of NVIDIA Quantum Switches:

1. High Bandwidth and Low Latency

• Quantum switches support speeds ranging from 40Gb/s (older models) up to 800Gb/s in the latest Quantum-X800 series.

2. In-Network Computing

• A standout feature is NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARP), which offloads collective communication operations (like data reductions in MPI) to the switch itself. This reduces network traffic and boosts application performance—SHARPv4 in the Quantum-X800 provides up to 14.4 teraflops of in-network computing power.

3. Scalability

• Quantum switches support massive topologies, such as Fat Tree, Dragonfly+, or multi-dimensional Torus. The Quantum-2 QM9700, for instance, can scale to over 2,000 nodes out of the box, while the Quantum-X800 Q3400 supports up to 10,368 nodes in a two-level fat-tree setup.

4. Advanced Features

• Adaptive Routing: Dynamically optimizes data paths to avoid congestion.

• Congestion Control: Ensures consistent performance even under heavy loads.

• Self-Healing Interconnect (SHIELD): Recovers from link failures 5,000x faster than software-based solutions, enhancing reliability.

• Performance Isolation: Enables multi-tenant environments by isolating workloads, crucial for cloud providers.

5. Switch Models

• Quantum-1: Earlier generation with speeds like 100Gb/s or 200Gb/s (such as QM8700).

• Quantum-2: Introduced 400Gb/s ports (such as QM9700 with 64 ports at 400Gb/s or 128 at 200Gb/s), delivering 51.2Tb/s bidirectional throughput.

• Quantum-X800: The latest (as of March 2025), with 800Gb/s ports (such as Q3400 with 144 ports in a 4U form factor or Q3200 with 72 ports in 2U), doubling bandwidth and targeting trillion-parameter AI models.

Applications:

• AI and Machine Learning: Powers GPU clusters (such as NVIDIA DGX systems) for training trillion-parameter models, as seen in xAI’s Colossus supercomputer.

• HPC: Accelerates simulations in physics, climate modeling, and genomics.

• Cloud-Native Supercomputing: Supports multi-tenant, bare-metal performance with security and isolation, used by providers offering HPC-as-a-service.