CUDA/C++ Performance Engineer for Differentiable Physics Simulator

• Profile bottlenecks across CUDA kernels, memory transfers, synchronization, sparse assembly, solver steps, and differentiable rollout paths.
• Determine the highest-impact performance lever for each bottleneck, whether kernel tuning, data residency, batching, stream usage, solver changes, or reduction/assembly redesign.
• Improve existing CUDA backend architecture, including host/device data flow, CUDA kernels, sparse assembly, and solver structure.
• Evaluate tradeoffs between targeted optimization, architectural refactoring, and larger rewrites when justified by profiling evidence.
• Measure and validate improvements in speed, GPU utilization, correctness, and numerical reproducibility.
• Deliver results in accordance with project timelines.
• Prepare written and oral technical reports and demonstrations.
• Collaborate with our teams of scientists and engineers in Honda’s regional and global R&D offices. Communicate profiling results, tradeoffs, and implementation outcomes to audiences with varying CUDA experience.

Honda Research Institute USA (HRI-US) is seeking a self-motivated engineer to join our Intelligent Robotics Research division. This individual will improve performance of a CUDA/C++ differentiable physics simulator across the GPU backend, including CUDA kernels, host/device data flow, sparse solver structure, and backward workflows used in optimization. The work will require profiling forward and backward simulation workloads, identifying bottlenecks, improving GPU utilization, and reducing CPU/GPU synchronization and transfer overhead., * Strong C++ and CUDA C++ experience in production or research codebases.

• Proven experience profiling and optimizing CUDA kernels with tools such as NVIDIA Nsight Systems, Nsight Compute, or equivalent GPU profiling workflows.
• Comfortable editing low-level GPU code involving reductions, atomics, sparse matrices, memory coalescing, launch configuration, and synchronization.
• Experience reducing CPU/GPU transfer overhead using better data residency, batching, pinned memory, async copies, streams, or kernel fusion.
• Familiarity with numerical simulation, optimization, or differentiable physics workflows.
• At least 1 year of hands-on experience with the qualifications above.

Bonus Qualifications

• Experience with contact-rich physics simulation.
• Experience with differentiable simulation or trajectory optimization.
• Experience optimizing sparse linear solvers on GPU.
• Experience tuning CUDA MPS workloads or multi-process GPU scheduling.
• 3+ years of hands-on experience with the qualifications above preferred.