NVIDIA
Senior System Software Engineer - GPU Performance
$152K–$288K
US, CA, Santa Clara +1 · Senior · Full-time · Posted Oct 9, 2026
- Location
- US, CA, Santa Clara
US, Remote - Experience
- 3+ years in parallel programming; relevant performance engineering and HPC experience required, with no duration specified.
- Education
- M.S. (or equivalent experience) or PhD in Computer Science or a related field.
About the job
Analyzes and improves communication-library performance across large multi-GPU, multi-node HPC clusters, diagnosing issues across hardware and software stacks.
What you'll do
- Characterize performance, evaluate solutions, and analyze interactions between communication libraries and system components.
- Triage and identify root causes of customer performance issues.
- Build tools and infrastructure to visualize and analyze performance data.
What they ask for
- M.S. (or equivalent experience) or PhD in Computer Science or a related field, with relevant performance engineering and HPC experience.
- 3+ years of parallel programming experience and experience with at least one communication runtime (MPI, NCCL, UCX, NVSHMEM).
- Experience benchmarking and triaging performance on large-scale HPC clusters; ability to debug issues across the hardware/software stack.
- Implement micro-benchmarks in C/C++; proficiency in a scripting language, preferably Python.
Nice to have
- Practical experience with Infiniband/Ethernet networking, including RDMA, topologies, or congestion control; or debugging network issues in large-scale deployments.
- Familiarity with CUDA programming and/or GPUs, or Deep Learning Frameworks such as PyTorch and TensorFlow.
Pay
152,000 USD - 241,500 USD base salary for Level 3; 184,000 USD - 287,500 USD base salary for Level 4
Tools
MPI, NCCL, UCX, NVSHMEM, C/C++, Python, Kubernetes, SLURM
AI focus
The role supports communication libraries used in deep learning and HPC applications. NVIDIA states that it uses AI tools in its recruiting processes.
Model-written summary · the original posting is authoritative