NVIDIA
Principal Deep Learning Communication Architect
$272K–$431K
US, CA, Santa Clara +2 · Principal · Full-time · Posted Apr 14, 2026
- Location
- US, CA, Santa Clara
US, TX, Austin
US, Remote - Experience
- 12+ years of industry experience in high-performance computing (HPC) or distributed deep learning.
- Education
- Ph.D. or M.S. in Computer Science, Electrical Engineering, or a related field (or equivalent experience).
About the job
Leads the architecture and development of communication libraries and algorithms that scale AI and HPC workloads across large clusters and next-generation NVIDIA platforms.
What you'll do
- Define the technical roadmap for communication libraries and scalable cluster architectures.
- Lead design of communication primitives and collective algorithms for heterogeneous interconnects.
- Co-design libraries and hardware with application developers, silicon architects, and software engineers; develop models and simulators to predict system behavior.
What they ask for
- Ph.D. or M.S. in Computer Science, Electrical Engineering, or a related field (or equivalent experience).
- Deep understanding of 3D parallelism and advanced strategies including Context Parallelism, Expert Parallelism, and ZeRO variants.
- Technical proficiency with NCCL, UCX, UCC, NVSHMEM, or MPI; experience with RDMA, RoCE, and low-level InfiniBand verbs is required.
- Advanced knowledge of inference engines and schedulers, and expert knowledge of NVIDIA GPU memory hierarchy and CUDA programming models.
Nice to have
- Hands-on experience with Megatron-Core, DeepSpeed, or JAX/XLA; upstream open-source contributions.
- Experience deploying and optimizing models on NVIDIA or similar rack-scale systems; patents or papers in top-tier systems or architecture venues.
Pay
272,000 USD - 431,250 USD (base salary)
Tools
NCCL, NIXL, NVSHMEM, UCC, UCX, MPI, RDMA, RoCE
AI focus
Build communication systems for trillion-parameter models and Agentic AI, including co-design for AI workloads and high-throughput inference.
Model-written summary · the original posting is authoritative