NVIDIA is looking to develop and productize deep learning solutions in autonomous driving vehicles by developing compiler technology to optimize larger and better models to leverage NVIDIA's unique hardware architecture.
Requirements
- Strong Python and/or C/C++ programming skills
- Proven technical foundation in CPU and GPU architectures, containers (nvidia-docker), numeric libraries, modular software design
- 2+ years of experience in implementing deep learning models and optimizations such as graph fusions, kernel implementation, KV Caching etc.
- Domain experience in current innovative deep learning methods (e.g. diffusion models, vision language action models, etc.)
- Experience with NVIDIA software libraries such as CUDA and TensorRT
- In depth experience with the internals of deep learning frameworks such as PyTorch or JAX (e.g. Creating custom operators, Graph fusions, deployment of PyTorch models)
Responsibilities
- Developing compiler technologies to run various classes of model architecture (Transformer, Diffusion, VLA, CNN, RNN etc.) on NVIDIA hardware leveraging techniques such as reduced precision, quantization, workload scheduling and memory bandwidth optimization
- Working across the whole lifetime of a model: training, fine-tuning, optimization to allow customers to access pioneering models on NVIDIA hardware
- Develop workflows that let users leverage frameworks (e.g. PyTorch, JAX) and ecosystem tools (HuggingFace, MLIR) without forgoing performance
- Stay up to date with the latest research and innovations in deep learning, implement and experiment with new insights to improve NVIDIA's Physical AI DNNs.
Other
- MS or PhD degree in computer science, computer vision, robotics, computer architecture or equivalent experience in technical field
- 5+ years of work experience in software development.
- Willing to take action and have strong analytical skills.
- Strong time-management and organization skills for coordinating multiple initiatives, priorities and implementations of new technology and products into very sophisticated projects.
- Background with low precision inference, quantization, compression of DNNs
