NVIDIA’s NVLink 6 keeps AI factories running through any error

NVIDIA’s new Vera Rubin NVL72 platform connects 72 Rubin GPUs as a single compute unit, delivering a substantial increase in density for AI infrastructure. The system is designed with uncompromising reliability in mind; according to NVIDIA, “A single dropped packet cannot be allowed to spike inference latency or disrupt a training collective.” NVLink 6 achieves this through a multi-layered resiliency framework, spanning hardware, system design, and software, that natively detects, contains, and recovers from errors to maximize continuous output for large-scale AI factories. This approach enables NVIDIA to offer a platform capable of training with ¼ the GPUs while maintaining high inference throughput. NVIDIA’s Vera Rubin NVL72 platform establishes a single scale-up domain by connecting 72 Rubin GPUs, a density increase that helps maintain output in large AI deployments. This architecture relies on the newly implemented NVLink 6 networking fabric to ensure these GPUs function as a unified compute unit, demanding a level of reliability previously unattainable at this scale. The system’s design prioritizes uninterrupted operation, recognizing that even infrequent packet loss can severely degrade performance within an AI factory environment. Traditional approaches to network reliability, such as Priority Flow Control and Explicit Congestion Notification, often introduce vulnerabilities like head-of-line…