A 311 model farm needs RDMA over Converged Ethernet between GPU workers on Spectrum-X. Which action aligns with that requirement?
Select an answer to reveal the explanation.
Short Explanation
The 311 farm wants GPU workers to pass tensors like firefighters passing hose—fast and direct. Spectrum-X RoCE is how RDMA rides the Ethernet fabric between those workers instead of dragging everything through slow detours.
Full Explanation
A 311 model farm that needs RDMA over Converged Ethernet between GPU workers on Spectrum-X should apply Spectrum-X RoCE configuration so those workers can use RDMA over the Ethernet fabric. RoCE is the mechanism that carries RDMA on Spectrum-X Ethernet; enabling and aligning it on the worker-facing AI path lets 311 models exchange gradients and activations with low latency while staying on the municipal Ethernet design. Enabling Spectrum-X InfiniBand transport so GPU workers can use native IB queue pairs is incompatible with the requirement because Spectrum-X delivers RDMA over Converged Ethernet, not native InfiniBand, and queue pairs are an IB concept rather than an Ethernet RoCE configuration. Configuring a Spectrum-X VXLAN overlay so GPU workers can stretch Layer 2 across racks adds encapsulation and tunneling complexity instead of providing the lossless, low-latency RDMA path that RoCE delivers directly over the Spectrum-X Ethernet fabric. Applying Spectrum-X adaptive routing configuration so GPU workers can balance flows across uplinks does not itself enable RDMA over Converged Ethernet; the requirement is RoCE on the worker-facing AI path, not merely load balancing across links. Exam caveat: “RDMA over Converged Ethernet” on Spectrum-X maps directly to RoCE configuration on the AI worker path—not InfiniBand transport with native IB queue pairs, a VXLAN Layer 2 overlay across racks, or adaptive routing that merely balances flows across uplinks. Operational check: apply RoCE on Spectrum-X interfaces serving 311 GPU workers, align host settings, confirm RDMA connectivity between workers, and ensure worker-facing ports remain up and dedicated to the AI fabric rather than enabled for InfiniBand transport, encapsulated in a VXLAN overlay, or left to adaptive routing across uplinks.