Workload-classified GPU fleet telemetry

FARMGPUbuyer room · privateManage buyer room

Workload-classified GPU fleet telemetry: a catalog of 50,000+ distinct GPU workloads over a 90-day window, each a ~100-field profile joining hardware telemetry (utilisation, memory, power draw, temperature, NVLink, ECC error counts) with container-level metrics and an LLM-assigned semantic workload label derived from image name, metric shape, and GitHub/Docker Hub repo lookup. Uniquely, it pairs measured utilisation with what is actually running and why — a demand-composition and efficiency asset rather than another price feed — plus a quantified waste layer (GPU-hours burned below 10 percent utilisation, worth $500K+ annually on the reference fleet) and an observed workload-mix distribution across development, image generation, inference, video generation, fine-tuning and training.