nemo-mbridge-perf-moe-hardware-configs

This NVIDIA skill is a hardware-oriented reference of representative Mixture-of-Experts (MoE) training playbooks for Megatron Bridge. It solves the planning problem of choosing a starting parallelism strategy and setting throughput expectations for a given GPU platform and model family, so engineers do not have to guess dispatcher, tensor/expert/pipeline parallel degrees, or realistic performance bands from scratch.

The reference is organized as quick-answer tables and config families. A platform playbook maps H100, B200, GB200, and GB300 to their typical MoE strategy (DeepEP vs HybridEP, precision like MXFP8, CUDA-graph and CPU-cleanup considerations) and what matters most on each. A first-answer checklist gives canonical parallelism layouts for workloads such as DSV3 and Qwen3 235B, deliberately marking unknown values as unspecified rather than extrapolating. Rounded performance bands express throughput as durable planning ranges (TFLOPS/GPU and MFU) rather than exact promises, and representative config blocks spell out dispatcher, TP/EP/PP/VPP, precision, recompute, and priorities per workload and platform. Cross-cutting notes cover pipeline-parallel layout notation and environment tuning knobs, with an explicit warning not to cargo-cult tracker rows.

It targets performance engineers, ML infrastructure teams, and researchers sizing or tuning large MoE training runs who want a trustworthy, platform-aware starting point. The content is purely informational configuration guidance with no scripts that modify a system; a static-validation note flags an international phone number present in a card file. Its published benchmark reports a passing evaluation with a 100% security dimension.