Principal Deep Learning Communication Architect
Core
Define long-term technical roadmap and design next-generation communication primitives and collective algorithms for NVIDIA's GPU platforms to enable scaling of trillion-parameter AI models across hundreds of thousands of nodes.
Role type
Principal Deep Learning Communication Architect
Builds
Next-generation communication libraries (NCCL, NVSHMEM, UCC, UCX) and collective algorithms for heterogeneous interconnects (NVLink, Spectrum-X, Quantum-X)
Domain
High-performance computing (HPC) and distributed deep learning
Deliverable
production ML models
Required skills
3D parallelism (Data, Tensor, Pipeline, Context, Expert, ZeRO), RDMA/RoCE/InfiniBand verbs, GPU memory hierarchy (HBM3e/HBM4), CUDA programming, high-throughput inference engines (TensorRT-LLM, vLLM, SGLang, Dynamo)
Preferred skills
Framework development (Megatron-Core, DeepSpeed, JAX/XLA), upstream open-source contributions, rack-scale system deployment, patents/papers in systems/architecture venues
Technologies
NCCL, UCX, UCC, NVSHMEM, MPI, NVLink, Spectrum-X, Quantum-X, TensorRT-LLM, vLLM, SGLang, Dynamo, Megatron-Core, DeepSpeed, JAX, XLA, PyTorch Distributed, KServe, Ray
Responsibilities
Define long-term technical roadmap for communication libraries; Lead development of next-generation communication primitives and collective algorithms; Partner with application developers to architect specialized communication primitives; Collaborate with silicon architects to influence hardware specifications; Develop high-fidelity analytical models and simulators
Seniority
Principal, hands-on IC
