Software Development · Santa Clara, CA

Full-time

Software Engineer, ML Performance

NVIDIA · Santa Clara, CA · Posted Nov 19, 2026

This posting comes from NVIDIA. Confirm the details on their site before applying.

Overview

NVIDIA has been transforming computer graphics, PC gaming, and accelerated computing for more than 25 years. It’s a unique legacy of innovation that’s fueled by great technology—and amazing people. Today, we’re tapping into the unlimited potential of AI to define the next era of computing. An era in which our GPU acts as the brains of computers, robots, and self-driving cars that can understand the world.

Responsibilities

  • Analyze the performance of distributed training and inference workloads throughout the entire stack: model and framework, data loading and checkpointing, compiler and kernels, CUDA and NCCL, host and network.
  • Define the performance metrics tracked across a fleet of GPUs. Break them down into factors engineers need to debug. Decide which signals the platform collects first.
  • Drive the analysis of regressions: segment workloads across machines, versions, and configurations to isolate the cause, and turn the method into analysis others can reuse.
  • Evaluate automated diagnoses: judge whether a finding is accurate, identify misleading results and missed problems, and build the labeled cases that keep the analysis honest.
  • Work closely with build partners on their real workloads, carry findings through to improvements, and feed what the product is missing back into the roadmap.
  • Bachelor’s degree in Computer Science, Computer Engineering, or a related field, or equivalent experience.
  • 12+ years of relevant software engineering experience, including hands-on optimization of GPU training or inference workloads in production.
  • Demonstrated performance work across more than one layer of the stack, with a clear account of the baseline, evidence, alternative explanations and validated outcome.
  • Experience diagnosing performance across many jobs, machines or tenants, not only tuning a single workload.
  • Enough familiarity with framework, runtime and compiler interfaces to say where a signal should be measured, even when someone else writes the instrumentation.
  • Ability to distinguish symptoms from causes, recognize noisy or incomplete measurements and the effects of the measurement itself, and know when the evidence is insufficient.
  • Strong programming and debugging skills, demonstrated by maintainable analysis code, tests or repeatable investigation procedures, including when using automated development tools.
  • Clear communication of findings, trade-offs and remaining uncertainty to the people who own the workload.
  • Proficiency with Python and CUDA.

Preferred qualifications

  • Improved operation of both training and inference.
  • Deep knowledge of one or more of PyTorch, JAX, Megatron, NCCL, TensorRT-LLM, vLLM, SGLang or Dynamo.
  • Experience moving between application, system software, driver or hardware roles, and the full-stack view that comes with it.
  • Statistical judgment applied to noisy, multi-node fleet-scale performance data. Ideally, experience with counter-factual analysis and segmentation of workloads for diagnosing regressions.
  • Built debugging or performance-analysis tools used by other engineers, with evidence that they made investigations more effective.

Great fits start here.

Whether you're looking for your next role or your next team member, give people and possibilities a place to meet.