4. Real-World Use Cases

Use Performance Explorer to gain actionable insights and streamline AI deployments:

• Make Data-Driven Decisions About Cluster Scale

  • Use case.

    Evaluate AI workload performance at various GPU counts to balance training time and cost.

  • Key actions.

    Examine the tradeoffs between Total Cost to Train and Total Time to Train across cluster scales to determine the optimal GPU count for your workload.

  • Example scenario.

    Examine the performance scaling characteristics of the NVIDIA DGX Cloud H100 Reference Architecture for your workload to quantify how adding GPUs impacts total cost.

• Navigate the Precision Switch From BF16 to FP8

  • Use case.

    Decide whether to switch from BF16 to FP8 precision.

  • Key actions.

    Compare how BF16 and FP8 precision affect training throughput and project costs.

  • Example scenario.

    Use DGX Cloud Benchmarking to compare FP8 and BF16 performance for your workload.

• Assess Performance Gains From the Latest NeMo Framework

  • Use case.

    Determine if upgrading to the latest NVIDIA NeMo Framework version improves your AI workload performance.

  • Key actions.

    Measure performance metrics from your current framework version with the latest version.

  • Example scenario.

    Evaluate how recent NeMo Framework updates affect your workload’s training throughput and potential cost savings.