How GPUVerse thinks
Workload, constraint, candidate, evidence, decision. The five nouns the whole system is built from.
The shared vocabulary of GPUVerse, the entities every recommendation is built from.
Everything begins with a workload
The domain model is deliberately provider-, hardware-, and cloud-agnostic: Organization → Project → Workload → Recommendation → Deployment → Observability → Optimization.The primitives
Workload
The central object, the computational task you want to execute (inference, training, batch, fine-tuning, embeddings, evaluation). Every recommendation begins with a workload.
Recommendation
The core output, a provider, compute resource, region, estimated cost, performance, risk, an explanation, and a confidence score. Every recommendation must be explainable, and is immutable once written.
Plan (Architecture)
A complete infrastructure design for a workload, compute, networking, storage, scaling, deployment, security, and monitoring, with a full cost breakdown. The planner reuses the recommendation engine.
Provider
A compute source, AWS, CoreWeave, Runpod, Lambda, Azure, Google Cloud, and others. GPUVerse treats every provider as interchangeable infrastructure.
Compute Resource (GPU)
The hardware executing workloads, GPU, TPU, NPU, CPU, and future accelerators, abstracted behind a common interface.
Confidence & Explainability
Every recommendation carries a confidence score (0.0–1.0) and human-readable explanations, following the principle “explain before you automate.”
Was this page helpful?