Understanding AI Unit Consumption And How To Optimize Usage?

News Knowledge Base
1

How to understand AI Unit consumption and how to optimize usage?

What are AI Units, and why are they needed for AI Center and/or Document Understanding?

AI Units is the measure used to license UiPath AI products such as AI Center, Document Understanding, Communications Mining, and Task Mining. For this article, the focus will be regarding the usage of AI Units in relation to AI Center and Document Understanding. AI Units are charged based on consumption, but oftentimes, understanding how AI Units are being consumed and how usage can be optimized is a bit confusing. Please see the information below to gain a better understanding of how usage is calculated, possible reasons for spikes in consumption, and a few tips on how to optimize usage.

How is usage calculated?

To calculate the overall consumption cost, the following formula is used:

  • prediction cost + hardware cost = consumption cost

Prediction Cost

When it comes to prediction cost, it is again dependent on the following 2 factors

  1. Input size
  2. Model used

Hardware Cost (Not applicable for skills deployed On-Prem)

When it comes to Hardware Cost, this is calculated by the following formula:

  • replicas x resource = hardware cost

Review the following document for specific details about Hardware Cost .

Pipelines

Regarding pipelines, the cost of AI Units is calculated based on the configurations used during deployment.

  1. CPU : 6 AI Units / hour
  2. GPU : 20 AI Units / hour

What could be the reason for high consumption/usage spike?

ML Skills:

  • ML Skills deployed with GPU or an increased number of replicas and/or resources. To check this click on the ML Skill and click Modify current deployment, to see if GPU is enabled or if Advanced Infra settings have been enabled with an increased number of replicas and/or resources (CPU+RAM).
  • ML Skills making predictions for a high volume of documents
  • ML Skills making predictions for large documents
  • Using the Predict button many times to call a skill to make labeling predictions for documents in Document Manager
  • If ML Skills are deployed in UiPath Cloud AI Center, one common misconception is that AI Units are only consumed when a skill is being called to make predictions, however, it must also be taken into consideration that when an ML Skill is deployed and showing as Available, AI Units are being consumed due to the hardware required for the skills to be deployed and made available for consumption.

Pipelines

  • Auto retraining is turned on for a pipeline. To check this, click on a pipeline, click on parameters. A drop-down should show a list of parameters that have been set. A Boolean called auto_retraining will be set to true here if auto_retraining was configured when the pipeline was created
  • GPU is enabled for pipeline runs. To check if GPU was used when a pipeline was configured, click on the pipeline and look to see if GPU has an X or a checkmark. If a checkmark is visible, GPU was used which would cause more AI Units to be consumed per hour. (Note: If running a pipeline with a very large dataset, it may actually be beneficial to enable GPU as the overall training time (hours) would decrease thus less AI Units could possibly be used in the long run.)
  • Large datasets cause long-running pipelines. The larger a dataset is, the longer it will take to train a model.

Optimization Techniques:

  1. Stop ML Skills if they are not being used. (Alternatively, set a timeframe for a skill to undeploy after a period of inactivity. This setting can be found in the modify current deployment menu for the skill.)
  2. Turn off auto-retraining for pipelines
  3. Prune datasets to remove unnecessary training documents. See the guide for Training High Performing Models .


For additional information about AI Units, read UiPath AI Units FAQ document.