Soumyadip Ghosh
Solutions Architect at NVIDIA
California, United States
Summary
🎓
Top SchoolSoumyadip Ghosh is a Solutions Architect and AI software engineer with a PhD and eight years of experience optimizing LLMs and scientific HPC workloads across CPUs, GPUs and AI accelerators. He has driven production-grade LLM inference (LLaMA, Qwen) using vLLM, Kubernetes and RDMA-enabled techniques on Intel Gaudi and benchmarked them against NVIDIA systems for enterprise customers like IBM Cloud. His background in communication-efficient distributed ML and MPI/NCCL integration—developed during his PhD and LLNL NAS work—enabled ~70% communication reductions and scalable NAS on supercomputers. Soumyadip has optimized MLPerf-class scientific models (CosmoFlow, DeepCAM, AlphaFold2) with oneAPI and performance tools such as VTune and Intel Trace Analyzer, and he’s an active open-source contributor to repositories like pytorch/examples and LLNL/lbann. Based in California and now at NVIDIA, he combines deep research rigor with hands-on deployment experience, often translating low-level system tuning into customer-facing production guides.
7 years of coding experience
12 years of employment as a software developer
Bachelor of Electrical Engineering, Bachelor of Electrical Engineering at Jadavpur UniversityPh.D. Electrical Engineering, Ph.D. Electrical Engineering at University of Notre Dame
Stackoverflow
1reputation
0reached
0answers
0questions
Github Skills (11)
pytorch10
mpi10
machine-learning10
distributed-training10
ccl9
cuda9
c-language8
data-science8
cprogramming-language8
mnist7
deep-learning7
Programming languages (2)
Github contributions (5)
pytorch/examples
Aug 2020 - Jan 2021
A set of examples around pytorch in Vision, Text, Reinforcement Learning, etc.
Role in this project:
ML Engineer
ML EngineerContributions:18 reviews, 12 commits, 1 PR in 5 months
Contributions summary:Soumyadip primarily contributed to a distributed training example using PyTorch, focusing on model definition, data loading, and the implementation of distributed training with MPI. The user also included code for GPU usage with NCCL. The contributions showcase an understanding of distributed computing principles and model training. The user also made changes to remove hardcoded paths and formatting for readability.
pytorchvisiondeep-learningreinforcement-learningreinforcement
soumyadipghosh/eventgrad
Dec 2019 - Dec 2021
Contributions:41 commits, 18 pushes, 2 branches in 2 years
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.