I have been using the Document Understanding feature for invoice extraction and have set up an ML Skill with GPU enabled.
I understand that enabling GPU results in a consumption rate of 40 AI units per hour when the ML Skill is active. However, I have some doubts regarding the actual AI unit consumption:
Per-Page Consumption:
Is 1 AI unit consumed per PDF page processed? Or does the AI unit consumption depend on other factors like page size or content complexity?
Inconsistent Consumption:
On certain days, if the extraction process completes within 30 minutes, the AI units consumed seem to align correctly with the page count. However, this behavior is not consistent across all days, even when the process duration is similar.
Multi-Machine Execution:
When I run the extraction simultaneously on multiple machines, the AI unit consumption is noticeably higher compared to running it on a single machine.
Why does executing the process on more machines at the same time result in higher AI unit usage?
I’m a bit confused about how AI unit consumption is actually calculated in these scenarios. Could someone help clarify these points?
Per-Page Consumption:
The charging of AI units totally depends on type of no of pages, classifier & extractor you have used. Not the page size. Although there is a limit on how large document you can pass to DU, but it does not computed to AI unit.
Each time a classifier & extractor invoked, you have to pay AI unit no matter you have used a it for same page or not.
Inconsistent Consumption:
If the number of pages, used extractor & classifier same, process time is same, then the consumption of AI unit should be consistent.
Multi-Machine Execution:
The charging of AI unit is dependent on how much time the the extractor/classifier invoked.
Since you are using multi-machine execution, lets say 2 machine.
Then extractor/classifier will get invoked two times. Hence consumption will be larger.
API Overhead:
If you are using REST API, the digitization also charged in form of AI unit.
Additionally refer below link for understanding DU charging:
Please mark as solution, if it resolved your doubts
AI unit get consumed by number of times you call the service.
Each time you do the calssification/extraction, you are gonna spend the AI unit.
No matter even if it is a same document.
Eample: If I am processing 30 documents then for each document you have to pay AI unit.
If you have duplicate dicuement, the DU wont recognize it.
Kindly mark as solution, if it resolves your query.
Hey RN.
Keep in mind that when deploying your own skills, the cost comes from two sources: predicts (each time you infer a document) + hardware (each hour for which each replica of the skill is up).
Check this page for more information:
The hardware cost at the time of deploying ML Skills is calculated as follows:
