Nvidia HPC AI Software Architect 

Switzerland, Vaud 

325198767

time type

Full time

posted on

Posted 4 Days Ago

job requisition id

What you will be doing:

• Design and prototype scalable software systems that optimize distributed AI training and inference—focusing on throughput, latency, and memory efficiency.

• Develop and evaluate enhancements to communication libraries such as NCCL, UCX, and UCC, tailored to the unique demands of deep learning workloads.

• Collaborate with AI framework teams (e.g., TensorFlow, PyTorch, JAX) to improve integration, performance, and reliability of communication backends.

• Co-design hardware features (e.g., in GPUs, DPUs, or interconnects) that accelerate data movement and enable new capabilities for inference and model serving.

• Contribute to the evolution of runtime systems, communication libraries, and AI-specific protocol layers.

What we need to see:

• Ph.D. or equivalent industry experience in computer science, computer engineering, or a closely related field.

• 2+ years of experience in systems programming, parallel or distributed computing, or high-performance data movement.

• Strong programming background in C++, Python, and ideally CUDA or other GPU programming models.

• Practical experience with AI frameworks (e.g., PyTorch, TensorFlow) and familiarity with how they use communication libraries under the hood.

• Experience in designing or optimizing software for high-throughput, low-latency systems.

• Strong collaboration skills in a multi-national, interdisciplinary environment.

Ways to stand out from the crowd:

• Expertise with NCCL, Gloo, UCX, or similar libraries used in distributed AI workloads.

• Background in networking and communication protocols, RDMA, collective communications, or accelerator-aware networking.

• Deep understanding of large model training, inference serving at scale, and associated communication bottlenecks.

• Knowledge of quantization, tensor/activation fusion, or memory optimization for inference.

• Familiarity with infrastructure for deployment of LLMs or transformer-based models, including sharding, pipelining, or hybridparallelism.