On September 15, 2026, NVIDIA published a technical deep-dive explaining how NVLink 6 delivers a natively lossless fabric for large-scale AI factories. The sixth-generation interconnect combines lightweight Forward Error Correction (FEC), Physical Layer Retry (PLR), and Universal Physical Layer (UPHY) recovery to correct signal errors at the silicon level with near-zero latency penalty. Credit-based flow control (CBFC) eliminates packet loss by design at the Link Layer.

The post, authored by Farshad Ghodsian, Jesse Clayton, Jim Dinan, and Oren Reiss, describes a five-layer resiliency stack spanning the Physical Layer (FEC, near-instantaneous), Link Layer (UPHY recovery and PLR, under 1 ms), Link Management (access and trunk link rebalancing, CBFC), Application Layer (software recovery, ~500 ms), and System Layer (rack-level resiliency, over 1 minute). NVLink 6 ships with the Vera Rubin platform, where a Vera Rubin NVL72 rack connects 72 Rubin GPUs into a single scale-up domain at 3.6 TB/s bidirectional bandwidth per GPU and 260 TB/s aggregate rack-level bandwidth.

Confirmed

  • NVLink 6 uses a lightweight FEC architecture paired with PLR as a rapid second line of defense, avoiding heavy-weight FEC algorithms that impose multi-hop delays. The transmitting port appends advanced error-correcting codes; the receiver reconstructs corrupted bits inline, correcting single or multi-bit errors with near-zero latency penalty.
  • When an error burst exceeds FEC correction capabilities, PLR retransmits packets directly at the physical layer, reducing packet drops to zero without involving higher-level software stacks.
  • If severe degradation triggers a physical Link Down event, UPHY recovery rapidly recalibrates physical parameters while packets are held in a hardware replay buffer to ensure zero data loss.
  • Credit-based flow control (CBFC) guarantees lossless transmission at the hardware level: a sender never injects a packet unless it holds credits indicating the immediate next hop has buffer space. This proactive approach eliminates packet loss by design and avoids the head-of-line blocking, PFC storms, and deadlocks associated with Ethernet Priority Flow Control (PFC).
  • The Link Manager autonomously performs access link and trunk link rebalancing when underlying physical links degrade, containing hardware faults locally without cascading retransmissions or collective stalls.
  • The NMX Controller places affected links into a "contain and drain" state, enabling hardware to retrain degraded links without data corruption while preventing fabric-wide back-pressure.
  • NMX High Availability (NMX-HA) automatically migrates the functional controller to an alternate switch tray within seconds if the primary host fails. The NVLink switch tray data plane is decoupled from the management CPU running NVOS, so an unplanned CPU reset or OS failure does not interrupt data-plane forwarding.
  • According to NVIDIA's internal testing, Shadow Engine Recovery in NVIDIA Dynamo maintains a fully initialized, idle replica process alongside the active inference engine. On B200 GPUs, NVIDIA reports this reduces inference downtime from 283 seconds for a cold restart to 7.3 seconds.
  • NCCL prototype support for cuda-checkpoint enables multi-node checkpoints that capture nearly all startup work; general availability is expected by the end of 2026.
  • NVLink Fusion extends the same multi-layer resiliency stack to custom XPUs, allowing third-party silicon to inherit the NVLink scale-up fabric and fault tolerance. Every NVLink Fusion deployment must include at least one NVIDIA product (GPU, CPU, NVLink switch, or ConnectX NIC).

Unknown

  • Independent replication of the 7.3-second Shadow Engine Recovery figure on B200 GPUs; the benchmark comes from NVIDIA's internal testing.
  • Exact rollout timeline and pricing for NVLink Fusion integrations with third-party XPUs beyond the announced AWS Trainium4 partnership.
  • Whether the NCCL cuda-checkpoint GA target (end of 2026) will be met, and what fraction of startup work the checkpoints capture in production workloads.
  • Real-world mean-time-between-failures (MTBF) improvements for Vera Rubin NVL72 deployments versus prior-generation NVLink 5 systems.
  • Power and area overhead of the lightweight FEC + PLR + UPHY recovery stack at 3.6 TB/s per GPU signaling rates.

Our take

NVIDIA treats resiliency as a full-stack differentiator, not a checklist. Pushing error correction to the physical layer with lightweight FEC and PLR, coupled with credit-based flow control that mathematically eliminates packet drops, removes the failure modes that plague Ethernet-based scale-up fabrics. The 7.3-second recovery figure is a vendor benchmark on B200 hardware; the real test is independent validation on Vera Rubin and whether the cuda-checkpoint GA timeline holds.

Series: 1. Nvidia Commits Up to $3 Billion to Lancium, Powering the Stargate AI Campus in Texas · 2. NVIDIA details NVLink 6 multi-layer resiliency stack for AI factories · NVIDIA AI Factory

Sources