Description
We are seeking a software optimization engineer to join our numerics team at NVIDIA. You will work on techniques to make large language models (LLMs) more efficient, focusing on optimizing GPU usage.
Your primary responsibility will be to investigate and optimize AI workloads across GPU systems, identifying bottlenecks and implementing improvements to increase throughput and reduce GPU hours.
Key responsibilities:
- Investigate the performance of evaluation and other AI workloads across GPU systems
- Identify bottlenecks in computation, memory, communication, data movement, orchestration, and resource allocation
- Evaluate and improve parallelism configurations, including how work is distributed across GPUs and nodes
- Understand how effectively workloads use GPU capabilities, including tensor cores, and identify opportunities to improve utilization
- Implement and validate changes that reduce GPU hours per unit of useful work or increase throughput
- Design clear benchmarks and measurements, accounting for tradeoffs among throughput, latency, resource use, and the quality or correctness of results
- Partner with workload owners to understand requirements, test improvements, and help deploy solutions
- Independently prioritize and drive optimization projects from investigation through measurable results
Requirements:
- Experience analyzing and improving the performance of compute-intensive software
- Bachelor's degree or equivalent experience, and 5+ years of relevant experience
- Ability to reason about and optimize parallelism strategies – including tensor, data, pipeline, context, and expert parallelism – to balance throughput, communication overhead, and GPU resource use
- Investigate unfamiliar code/systems and work directly with code owners to land end-to-end optimizations
- Understanding of GPU architecture and the factors that affect utilization and application performance
- Experience using profiling, benchmarking, and measurement to guide optimization decisions
- Ability to work independently, communicate findings clearly, and collaborate with researchers and engineers
Preferred qualifications:
- Optimizing AI training, inference, evaluation, or other large-scale GPU workloads
- Working with GPU programming or machine learning frameworks and performance tools
- Improving workload scheduling, batching, data pipelines, or multi-GPU configurations
- Familiarity with collection and interpretation of GPU telemetry
- Energy-proportional computing or performance optimization under power, energy, or thermal design power budgets and experience with model quantization or numerical methods
In this role, success means measurably reducing the GPU resources required to deliver useful results across our workloads.
The expected base salary ranges for this position are: 150,000 USD – 286,000 USD for Level 3; and 183,750 USD – 351,000 USD for Level 4.
For more fresh new jobs, check out yubhub.co.