WORKLOAD GUIDE
Start with your workload.
Define the problem, measure a representative run, then select the infrastructure.
01 / REQUIREMENTS
Model training & fine-tuning
An accelerator specification is a starting point. A training plan also needs a model-memory strategy, data pipeline and tolerance for interrupted runs.
- Record parameter count, training precision, sequence length and target batch size.
- Separate model weights, optimiser state, activations and temporary memory.
- Measure storage throughput, checkpoint size and restart time on a representative run.
- Confirm topology, interconnect and scaling efficiency before committing to a larger allocation.
Illustrative application: A model team can compare an initial experiment with a longer dedicated run using its own rates and measured utilisation.
Plan a scenario →02 / REQUIREMENTS
Inference & agentic AI
The useful unit may be a request, token or completed task. GPU-hours alone do not predict response time or quality.
- Define target latency, peak concurrency and context length.
- Benchmark prefill and decoding separately using representative prompts.
- Identify CPU orchestration, tool execution, caching and retrieval dependencies.
- Compare reserved capacity with burst requirements, including cold starts and failure recovery.
Illustrative application: A product team can model the rental envelope for its base capacity and a separate peak scenario. It still needs workload benchmarks.
Plan a scenario →03 / REQUIREMENTS
Simulation & scientific research
HPC applications vary widely. Some scale across tightly connected CPUs; others use GPUs or independent task arrays.
- Confirm the application supports the chosen accelerator and software stack.
- Identify communication patterns, memory per task and job-scheduler needs.
- Account for software licensing, scratch storage and movement of large datasets.
- Retain reproducible environments, reference results and a documented validation method.
Illustrative application: An engineering team can compare a short burst of cloud capacity with a sustained research allocation without assuming linear speed-up.
Plan a scenario →Educational planning guidance. Use application documentation and measured benchmarks for an actual deployment.