300-640 — Frequently Asked Questions

While cisco_dlb=1 configures the NIC, the question asks for the switch configuration required to enable the integration. The switch must have the hardware profile enabled first to process the adaptive routing traffic.

It primarily affects traffic utilizing the Spectrum-X adaptive routing features. It enables the switch to understand and load balance based on the enhanced packet headers used by NVIDIA GPUs.

C-Series requires separate Fabric Interconnects and manual onsite setup, violating the 'no on-site IT' and 'remote launch' constraints.

Yes, it supports GPU nodes via the X-Fabric internal connectivity, enabling potential GPU acceleration for AI workloads.

SCU provides the necessary boot environment and drivers to automate the OS installation process on Cisco X-Series servers without user intervention.

Cisco Mode uses the SCU for automated deployment, while UEFI Mode typically relies on standard network boot protocols like iPXE to load the OS installer.

Inference serves live application requests, requiring immediate predictions from trained models to support user interactions or automated systems.

No, data preparation is still required to format and clean input data before it can be processed by the inference engine.

Speech recognition converts audio to text, while NLP interprets the meaning of that text to perform actions or extract structured data.

No, data compression is a standard IT infrastructure task for storage efficiency, not an AI-specific workload involving learning or inference.

Image recognition specifically targets visual data like photos and videos, whereas pattern recognition can apply to any data type including audio or numerical sequences.

No, anonymization is a data privacy technique. While AI can help blur faces, the core task of image recognition is identification and labeling, not privacy masking.

Aggregation of statistics is a monitoring and reporting function used to track system health, whereas scheduling actively assigns work to resources.

Backup generation is part of disaster recovery and data protection, while workload scheduling manages real-time task distribution across compute nodes.

No, RAG keeps the LLM's parameters frozen. It only adds retrieved context to the input prompt during inference.

No, RAG provides factual context but cannot fix inherent biases embedded in the base model's training data.

Model training involves adjusting weights and biases using training data to learn patterns. Deployment involves integrating the trained model into applications via APIs for inference.

Detecting model drift occurs after deployment to monitor performance degradation over time due to changing data patterns. Training focuses on initial learning from static datasets.

Isolation is important for security, but it does not inherently support the speed or scaling aspects of rapid provisioning required for dynamic AI workloads.

No, it applies to all infrastructure components including compute (GPUs/CPU) and storage, allowing them to be deployed quickly alongside network resources.

Option C describes restriction to local storage, which is the opposite of distribution. Workload distribution requires utilizing both local and external resources.

No, workload distribution focuses on task allocation and resource utilization. Security policies are a separate layer of infrastructure management.

RoCEv2 needs PFC to prevent packet loss during micro-bursts within the data center fabric (link-level). It needs ECN to signal end-hosts to throttle back when congestion persists or occurs at the endpoint (transport-level).

If you only use ECN, the network switches may drop packets due to full buffers before the end-hosts can react to the ECN marks, leading to poor performance and timeouts in RDMA applications.