01 / Inside the engagement
Situation
The company had built a GPU research platform on Azure Kubernetes for its scientists, but no team had proven a real workload on it. Outbound network access was locked down, containers could not be built inside the environment, and the first candidate was a demanding scientific simulation.
Before
- Scientific workload
- Standard environment
- Reference physics
After
- Scientific workload
- Azure AKS GPU pipeline
- Validated physics
Simplified workflow. The case text describes the contribution, delivery stage and evidence.
Explore the full conceptual diagram
02 / Inside the engagement
What UTIS did
- Built the first end-to-end machine-learned simulation pipeline on the platform (NVIDIA ALCHEMI and NIM, PyTorch, MACE), reproduced the known physics on H100 GPUs and benchmarked it against the standard environment.
- Packaged the pipeline for the client's own researchers.
- Turned the lessons into an onboarding kit: an intake template, an architecture reference and a "code-as-intake" method that derives GPU, data and MLOps needs straight from a team's repository.
- Right-sized a second workload moving onto the platform.
03 / Inside the engagement
Results
- ~10× faster than the standard environment on the same physics, in a validated proof of concept.
- Onboarding kit adopted by every research team on the platform; a team now onboards in about two weeks.
- Second workload right-sized to 50–75% fewer GPUs.
Why it matters to you
Validating a real workload gives researchers a tested starting point for an existing GPU platform. In this engagement, the reusable onboarding kit supported onboarding in about two weeks; future workloads still need their own reference checks.
Stack
How we measured it: benchmark against the standard environment with the same inputs; the reference physics was reproduced before timing was compared.