NVIDIA Quantum-X800 InfiniBand Platform
End-to-end, 800Gb/s networking with the highest performance for large AI. The first end-to-end 800Gb/s network in the world. The NVIDIA Quantum-X800 platform, which consists of the NVIDIA Quantum-X800 InfiniBand switch, NVIDIA ConnectX-8 SuperNIC, and LinkX cables and transceivers, is the next version of the NVIDIA Quantum InfiniBand and is designed specifically for trillion-parameter-scale AI models.
A new horizon of AI innovation is made possible by the new platform, which supports sophisticated hardware-based In-Network Computing with Scalable Hierarchical Aggregate Reduction Protocol (SHARP) v4, adaptive routing, and telemetry-based congestion control.
Multi-site AI factories with new co-packaged silicon photonic networking switches that can scale to millions of GPUs.
What’s Inside the Platform
NVIDIA Quantum-X800 InfiniBand Switches
Each of the 144 ports on the NVIDIA Quantum-X800 InfiniBand switch has 800Gb/s connection. Performance isolation capabilities, adaptive routing, telemetry-based congestion control, hardware-based In-Network Computing with SHARP v4, and a dedicated port supporting the Unified Fabric Manager (UFM) are all included. Advanced power-efficiency capabilities including power profiling and low-power link status are also added by NVIDIA Quantum-X800 switches.
The NVIDIA Quantum-X800 switch offers enhanced performance and power efficiency to drastically cut down on energy expenses, scientific computing, and the time it takes to complete AI workloads.
By reducing the distance and number of connections between optics and electronics, quantum-X silicon photonics switches further lower latency and overall power consumption.
NVIDIA ConnectX-8 SuperNIC
The most recent advancements in sophisticated In-Network Computing are supported by the NVIDIA ConnectX-8 SuperNIC, which provides 800Gb/s connectivity with extremely low latency. It still offers quality of service, adaptive routing, congestion control, faster MPI hardware engines, and more, all built on the ConnectX architecture.
LinkX Cables and Transceivers
With connectorized transceivers, passive fibre cables, and linear active copper cables (LACCs), the NVIDIA Quantum-X800 platform connectivity choices with the NVIDIA LinkX interconnect portfolio offer the most flexibility for creating a desired network topology.
Capabilities of NVIDIA Quantum-X800 InfiniBand platform
A high-performance, end-to-end 800Gb/s networking solution made for massive-scale AI is the NVIDIA Quantum-X800 InfiniBand platform. It is the NVIDIA Quantum InfiniBand’s next iteration.
The NVIDIA Quantum-X800 InfiniBand platform’s primary capabilities for large-scale AI are as follows:
- It is specifically designed for AI models with trillions of parameters.
- It facilitates sophisticated hardware-based, In-Network Computing with SHARP v4, opening up new avenues for AI development.
- It has telemetry-based congestion control and adaptive routing.
- With new co-packaged silicon photonic networking switches, the platform can grow to multi-site AI factories and millions of GPUs.
- It provides improved performance and power efficiency to drastically cut down on energy expenses, scientific computing, and the time it takes to complete AI workloads.
- It has the ability to isolate performance.
Advantages
- 800Gb/s end-to-end connectivity per port is the highest performance networking available. This is essential for high-performance computing and demanding AI workloads.
- Support for Scalable Hierarchical Aggregate Reduction Protocol (SHARP) v4 in sophisticated hardware-based In-Network Computing. This feature speeds up workloads in scientific computing and artificial intelligence by enabling effective group operations.
- congestion control based on telemetry and adaptive routing. These characteristics maximise network stability and performance, particularly in large-scale settings.
- connectivity with extremely low latency. Distributed training of big AI models and other latency-sensitive high-performance computing applications require low latency.
- improved efficiency of power. Features like power profiling and low-power link state are incorporated into the platform, and Quantum-X silicon photonics switches are used to further lower overall power consumption. For big data centers where energy expenses are an issue, this is a huge advantage.
- Scalability to multi-site AI factories and millions of GPUs. This suggests that the platform is made for extremely large-scale implementations, most likely in data center and cloud environments that support substantial AI infrastructure.
- The ability to isolate performance. In shared or multi-tenant computing environments, this functionality is crucial for maintaining constant performance across various workloads.
- The Unified Fabric Manager (UFM) is supported via a specific port. For large-scale networks in data centres and high-performance computing environments to operate effectively, UFM is a tool for controlling and monitoring InfiniBand fabrics.
- The NVIDIA LinkX interconnect range, which includes connectorized transceivers with passive fibre cables and linear active copper cables (LACCs), offers the greatest flexibility for network topology construction. This enables modification according to the particular requirements of various computing environments, including data centres with diverse sizes and configurations.
In conclusion, the NVIDIA Quantum-X800 InfiniBand platform provides a high-bandwidth, low-latency, and scalable networking architecture for massive-scale AI applications combining In-Network Computing and silicon photonics.
