Transit planners schedule a multi-day GPU job to update a traffic-prediction model from last year’s sensor dumps. Which workload type are they running?

Transit planners scheduling a multi-day GPU job to update a traffic-prediction model from last year’s sensor dumps are running a training workload. Training is characterized by sustained, data-heavy accelerator use that revises model weights from historical examples. Batch inference that is a long-running, data-heavy GPU job that uses frozen weights is a different phase: it applies an already trained model to generate predictions without revising weights, so it does not update the traffic-prediction model described in the stem. Model evaluation that is a long-running, data-heavy GPU job that tests forecast error is a validation step: it checks predictions against known outcomes but does not itself revise the traffic-model weights from sensor dumps. Feature engineering that is a long-running, data-heavy GPU job that builds datasets is an upstream preprocessing step: it prepares inputs but does not perform the weight update described in the stem. Exam caveat: after training completes, the same municipal platform will usually run batch inference that uses frozen weights on the new traffic model; do not confuse the long training job with that scoring path when sizing or scheduling. Operational check: confirm the job updates checkpoints/weights, uses historical dumps as training data, and is scheduled on the training cluster with appropriate east-west and storage I/O—not on a batch inference run that is a long-running, data-heavy GPU job that uses frozen weights.