When you train a model using a Tabular Workflow, you are charged based on the cost of the infrastructure and the dependent services. When you make predictions with this model, you are charged based on the cost of the infrastructure.
The cost of the infrastructure depends on the following factors:
- The number of machines that you use. You can set associated parameters during model training, batch prediction, or online prediction.
- The type of machines that you use. You can set this parameter during model training, batch prediction, or online prediction.
- The length of time for which the machines are in use.
- If you are training a model or making batch predictions, this is a measure of the total processing time of the operation.
- If you are making online predictions, this is a measure of the time that your model is deployed to an endpoint.
Tabular Workflows runs multiple dependent services in your project on your behalf: Dataflow, BigQuery, Cloud Storage, Vertex AI Pipelines, Vertex AI Training. You will be charged by these services directly.
Examples of training cost calculation
Example 1: 110MB dataset in CSV format, trained for one hour with default hardware configuration.
The cost breakdown for the default workflow with Architecture Search and Training is as follows:
| Service | Cost |
|---|---|
| Dataflow example and stats generation | $2 (Dataflow ran 7 min) |
| Dataflow data and feature transformations | $3 (Dataflow ran 10 min) |
| Vertex AI Training | 0.8hr x $20 + 0.2hr x $20 + $3.3 SSD cost + pipeline container cost = $24 (48min tuning, 12min training) |
| Vertex AI Pipelines | 1 run x $0.03 = $0.03 |
| Total excluding model distillation | $27.03 |
Optionally, you can enable model distillation to reduce resulting model size. The cost breakdown is as follows:
| Service | Cost |
|---|---|
| Total excluding model distillation | $27.03 |
| Vertex AI Training for model distillation | $1 |
| Dataflow data, feature transformations for model distillation | $3 (Dataflow ran 10 min) |
| Batch prediction for model distillation | $7 |
| Total including model distillation | $38.03 |
Example 2: 1.84TB dataset in BigQuery, trained for 20 hours with hardware override.
The hardware configuration for this example is as follows:
| Hardware Configuration Name | Value |
|---|---|
| stats_and_example_gen_dataflow_machine_type | n1-standard-16 |
| stats_and_example_gen_dataflow_max_num_workers | 100 |
| stats_and_example_gen_dataflow_disk_size_gb | 40 |
| transform_dataflow_machine_type | n1-standard-16 |
| transform_dataflow_max_num_workers | 100 |
| transform_dataflow_disk_size_gb | 200 |
| distill_batch_predict_machine_type | n1-standard-2 |
| distill_batch_predict_starting_replica_count | 200 |
| distill_batch_predict_max_replica_count | 200 |
The cost breakdown for the default workflow with Architecture Search and Training is as follows:
| Service | Cost |
|---|---|
| Dataflow example and stats generation | $518 (Dataflow ran 6 hours) |
| Dataflow data, feature transformations | $471 (Dataflow ran 6 hours) |
| Vertex AI Training | 17hr x $20 + 3hr x $20 + $41.5 SSD cost + pipeline container cost = $555 (17 hours tuning, 3 hours training) |
| Vertex AI Pipelines | 1 run x $0.03 = $0.03 |
| Total | $1544.03 |