Hardware
NVIDIA NVLink: The Scale-Up Network for AI Factories
NVIDIA’s sixth‑generation NVLink is a purpose‑built scale‑up network that delivers up to 3.6 TB/s per GPU and 260 TB/s rack‑level bandwidth with 130 TFLOPS in‑network compute—outperforming off‑the‑she
NVIDIA’s sixth‑generation NVLink is a purpose‑built scale‑up network that delivers up to 3.6 TB/s per GPU and 260 TB/s rack‑level bandwidth with 130 TFLOPS in‑network compute—outperforming off‑the‑shelf Ethernet for large MoE and LLM workloads. By co‑designing hardware and software (NVIDIA Dynamo, TensorRT‑LLM, NCCL, NIXL), NVLink supports disaggregated inference, expert parallelism, and dynamic resource allocation, yielding a 50× increase in tokens per watt from Hopper to Blackwell. The platform offers robust resiliency (control‑plane resilience, hot‑swappable trays, dynamic routing, fine‑grained telemetry) and forward‑compatibility (NVLink‑C2C for CPU‑GPU coherence, NVLink Fusion for custom XPUs), ensuring dependable ROI for production AI factories.
Source: NVIDIA Developer | 2026-07-20