Deep Learning Kernel Software Performance Architect
At a glance
Mid-level AI Research role at NVIDIA. China · full-time.
Pay not stated AI Research salary
aimlroles summary, based on the employer's posting.
What you'll do
- Analyze GPU kernels and deep learning operations against baselines or projections, then tune performance gaps
- Trace slowdowns from reproduction through root-cause isolation, proposed fixes, and closure with owning teams
- Build Python automation for performance measurement and analysis, keeping scripts maintainable and reviewable
- Set up repeatable test workflows, including workload coverage, large-scale runs, rerun rules, and regression checks
- Coordinate with kernel, compiler, architecture, QA, and infrastructure teams on scalable performance checks
What you bring
- A master's, doctorate, or equivalent experience in computer science, computer engineering, applied math, or a related area
- At least two years using Python and C/C++ to read or debug performance-focused code
- Working knowledge of computer architecture, parallel programming, and reasoning about latency, throughput, and memory
- Experience with profiling, reliable measurement, reproducible results, and performance regression triage
Who this fits
This mid-level role suits someone who can combine kernel performance analysis with Python automation and work issues through cross-team decisions. The role is based in China.
From the employer
NVIDIA is seeking Software Performance Architects to optimize GPU kernel performance for state-of-the-art data-center platforms. We build automated, data-driven workflows to detect, explain, and prevent performance regressions across key deep learning workloads, partnering closely with kernel developers, compiler teams, infrastructure, and architecture/performance groups.
What you'll be doing:
Performance analysis, optimization and debugging
Build performance narratives using structured methodology: baselines, projections, controlled comparisons, and regression attribution.
With the methodologies, analyze performance of GPU-accelerated kernels and key deep learning building blocks, identify gaps with baselines or projections, then optimize the kernels' performance to fill the gaps.
Debug performance issues end-to-end: reproduce, isolate root causes, propose fixes or mitigation paths, and drive closure with the owning teams.
Automation + regression infrastructure (Python-heavy)
Develop and maintain Python-based automation for performance testing and analysis—using modern AI-assisted developer tools (e.g., Cursor/Claude Code/Copilot) to accelerate scripting while keeping code maintainable and reviewable.
Design and operate performance test workflows: coverage definition, test/workload generation, automated large-scale execution (CI/nightly/on-demand), rerun rules, and reproducibility standards.
Cross-team collaboration and operating model
Work with kernel developers and the compiler teams to ensure performance checks are practical, scalable, and aligned to release needs.
Work with chip architecture and modeling teams to solidify the performance methodology across chip architecture generations and common Deep Learning operators such as GEMM, Attention, MoE.
Partner with SWQA and infrastructure teams for execution at scale and reliable pipelines/dashboards.
Following general software engineering best practices including support for regression testing and CI/CD flows
What we need to see:
Masters or PhD degree or equivalent experience in Computer Science, Computer Engineering, Applied Math, or related field
Strong programming ability in Python plus C/C++ with 2+ working experience (performance-oriented code reading/debugging)
Solid fundamentals in computer architecture, parallel programming and performance reasoning (latency/throughput, memory hierarchy, parallelism) to be able to identify bottlenecks, optimize resource utilization, and improve throughput
Experience with performance analysis workflows: profiling, measurement methodology, reproducibility, and regression triage.
Comfortable working across teams and driving issues to decision/closure with clear communication
Ways to stand out from the crowd:
Experience with high-performance kernels or math libraries (e.g., GEMM/attention, CUTLASS-like concepts)
GPU programming/perf experience (CUDA or equivalent parallel programming)
Strong ML/DL workload understanding (training/inference shapes, precision modes, perf bottlenecks)
Familiarity with simulators/analytical modeling or performance characterization methodology