Power & Performance Management

Key Responsibilities

• Benchmark power consumption for various AI Kernels/models on 3P products.
• Benchmark power consumption for various AI Kernels/models on 1P.
• Setup Scripts/Instrumentation needed for Power sampling.
• Collect power measurement, Perf traces, etc., required for a deeper dive.
• Setup Scripts/Instrumentation needed for Power sampling on AMD ROCm SMI and Nvidia SMI devices for various kernels.
• Data collection: Power, voltage, frequency, DVFS, PMON, HBM Bandwidth, NCU traces, etc.
• Collect data for understanding DCIT (Rack/Blade level power management, average and peak power on Platform) that helps with analysis to understand where the bottlenecks are and scope for improvement.
• Extract PerfWatt metric for various ranges of sweep parameters (e.g., BMNK sweep for Gemm kernels, Bandwidth sweep for HBM stream tests, etc.).
• Extract VF curves for various operating frequencies supported by 3P systems.
• Maintain a dashboard of power for various tests to compare with 1P.
• Automate spreadsheet-based power models and integration with Hiper incorporating changes to sweep for various workload parameters enabling architecture what-ifs for various options.

Requirements & Tech Stack

• Familiarity with Power and Performance measurements techniques on 1P.
• Familiarity with Nvidia SMI on H100, H200.
• Familiarity with AMD SMI on MI300 GPUs.
• Familiarity with Python, Cuda programming.
• Familiarity with Automation scripting in PowerShell, Python.
• Familiarity with Power BI Dashboards.