GPU Performance Engineer - Neural Reconstruction

NVIDIA•St. Louis, MO

1d•$224,000 - $431,250•Onsite

About The Position

NVIDIA is building the future of computer graphics, simulation, robotics, and embodied AI. Neural reconstruction and Gaussian Splatting are changing how 3D worlds are collected, represented, optimized, and rendered. These workloads push the limits of GPU computing, differentiable rendering, computer vision, and production ML systems. In this role, you will help make neural reconstruction faster, more scalable, and more reliable. You will work across PyTorch, CUDA, C++, and GPU profiling to optimize training and rendering workflows used in sophisticated 3D reconstruction systems. The ideal candidate enjoys working close to the hardware while understanding the ML and 3D vision goals behind the system.

Requirements

BS, MS, PhD, or equivalent experience in Computer Science, Computer Engineering, Electrical Engineering, Applied Math, Robotics, Computer Vision, Machine Learning, or a related field (or equivalent experience) with 12+ years of experience.
Strong programming skills in Python and C++!
Hands-on experience with PyTorch or a similar tensor/autograd framework.
Experience optimizing GPU-accelerated workloads using CUDA, C++/CUDA extensions, or related GPU programming approaches.
Practical experience with profiling and performance analysis, including root-causing CPU/GPU bottlenecks, synchronization overhead, memory pressure, kernel launch overhead, and framework-level inefficiencies.
Ability to develop benchmarks and validate that optimizations preserve correctness, numerical behavior, and user-visible quality.
Strong communication skills, including the ability to explain performance tradeoffs, risks, and results to research and engineering partners.

Nice To Haves

Experience with Gaussian Splatting, NeRF, differentiable rendering, rasterization, neural rendering, SLAM, 3D reconstruction, or robotics/autonomous-vehicle perception pipelines.
Deep CUDA performance experience, including memory access patterns, shared memory, atomics, occupancy, launch configuration, synchronization, and numerical stability.
Experience optimizing PyTorch workloads with custom operators, fused kernels, sparse tensors, distributed training, or distributed rendering.
Familiarity with camera and lidar geometry, projection models, calibration, rolling shutter, depth rendering, or multi-sensor reconstruction.
Experience improving large production ML systems where quality metrics, training speed, memory footprint, and developer velocity must be balanced.

Responsibilities

Profile end-to-end neural reconstruction workflows and identify bottlenecks across data loading, initialization, training, rendering, evaluation, and export.
Improve CUDA and PyTorch performance for Gaussian Splatting and neural reconstruction workloads, including camera/lidar data, multiview batching, large-scene rendering, and memory-sensitive training paths.
Analyze GPU performance using tools such as Nsight Systems, Nsight Compute, NVTX, PyTorch Profiler, CUDA events, and benchmark dashboards.
Optimize sparse and irregular rendering workloads, including tile-level masking/culling, sparse gradients, batching, and multi-GPU execution.
Translate high-impact Python, NumPy, or PyTorch bottlenecks into efficient CUDA/C++ or PyTorch-native implementations when appropriate.
Validate that performance improvements preserve reconstruction quality, numerical behavior, camera/lidar correctness, and production reliability.
Build repeatable benchmarks, regression tests, and profiling workflows to catch performance and quality regressions early.
Collaborate with researchers, CUDA engineers, ML engineers, and production teams to turn promising prototypes into maintainable, reviewable, production-quality code.