Understanding Capacity Units (CU)

[tan_thiamhuat]

can someone share how this Capacity Units are being computed when we run code in Fabric?

when running the same / identical feature, e.g copy dataset from Source to Destination, does the Notebook code use less CU, than the other low-code of Data Flow Gen 2? estimated how much percentage less?

Can you explain the concepts of Bursting and Smoothing?

If we buy a reservation to lock in 40% discount for a year, can we change the capacity during the 1 year, if we find the capacity is not enough? or the other way, can we downgrade if we find it is too much?

[burakkaragoz · Accepted solution]

Hi @tan_thiamhuat ,

Great set of questions — here's a breakdown based on what we know so far:

1. How are Capacity Units (CU) computed?

CU usage is based on the compute resources consumed during execution. This includes CPU, memory, and I/O operations. The more intensive your operation (e.g., large data transformations, joins, or writes), the more CU it consumes.

2. Notebook vs Dataflow Gen2 CU usage

In general:

• Notebooks tend to be more efficient in CU usage because they give you more control over execution (e.g., batching, filtering early, caching).
• Dataflow Gen2 is low-code and easier to use, but it may abstract away optimizations, leading to slightly higher CU usage for the same task.

There’s no official percentage published, but in practice, Notebooks can use 20–40% less CU depending on how optimized your code is.

3. Bursting and Smoothing

• Bursting allows workloads to temporarily exceed the provisioned CU limit by borrowing from unused capacity (if available). This helps avoid throttling during short spikes.
• Smoothing spreads out CU usage over time to avoid sudden spikes. It’s useful for workloads that have uneven demand patterns.

These mechanisms help maintain performance without over-provisioning.

4. Changing reserved capacity

If you purchase a 1-year reservation with a 40% discount:

• You cannot downgrade the capacity during the term.
• You also cannot upgrade mid-term. You’d need to purchase additional capacity separately.
• Microsoft recommends sizing carefully before committing to a reservation.

If flexibility is important, consider using pay-as-you-go until you’re confident about your workload patterns.

Hope this helps clarify things.

If my response resolved your query, kindly mark it as the Accepted Solution to assist others. Additionally, I would be grateful for a 'Kudos' if you found my response helpful.

[suparnababu8 · Accepted solution]

Hi @tan_thiamhuat 

Capacity Units (CUs) in fabric calculated based on your purchased SKU capacity. Computation is measured  in 30seconds. Pls have a look on below , how consumation is calculated.

While coming to your 2nd query, yes you are correct, Notebooks are consume less CU's as compared to Dataflow gen2. Dataflow gen2 supports for low-code development with built in transformaions, so it will consume more CU's. Microsoft Fabric – Comparing Dataflow Gen2 vs Notebook on Costs and usability - FourMoo | Microsoft ... as this blog says, notebook executon was around 10to20% high in CU's consumation as compared to dat flowgen2 for same workloads

While coming to Bursting concept, it allows the workloads to excced their allocated CU's limit. as I attended one seesion previous week with Micrsoft employee, In that session he explained for example if your capacity is an F8, and your workload takes 100 seconds to complete, and it uses 1500 CU, the scale factor would be calculated as follows: 1500 / 100 / 8 = 1.875. Pls see below image for SKU guardrils for burstable capacity. Read more detail hereBurstable Capacity - Microsoft Fabric | Microsoft Learn 

While coming to your last question, no you ahve to purchase a new capacity and you can't downgrade your reserved capacity until it completes one year. Read more details here Save costs with Microsoft Fabric Capacity reservations - Microsoft Cost Management | Microsoft Learn

Thank you!

Did I answer your question? Mark my post as a solution!

Proud to be a Super User!

[suparnababu8 · Accepted solution]

Hi @tan_thiamhuat 

It means that, 10to20% more efficent in CU's consumption. 

Here if you see notebooks consumption of CU's is low as compared to Dataflow gen2. read more about it here Microsoft Fabric – Comparing Dataflow Gen2 vs Notebook on Costs and usability - FourMoo | Microsoft ...

Thank you!

Did I answer your question? Mark my post as a solution!

Proud to be a Super User!

[suparnababu8 · Accepted solution]

Hi @tan_thiamhuat 

Capacity Units (CUs) in fabric calculated based on your purchased SKU capacity. Computation is measured  in 30seconds. Pls have a look on below , how consumation is calculated.

While coming to your 2nd query, yes you are correct, Notebooks are consume less CU's as compared to Dataflow gen2. Dataflow gen2 supports for low-code development with built in transformaions, so it will consume more CU's. Microsoft Fabric – Comparing Dataflow Gen2 vs Notebook on Costs and usability - FourMoo | Microsoft ... as this blog says, notebook executon was around 10to20% high in CU's consumation as compared to dat flowgen2 for same workloads

While coming to Bursting concept, it allows the workloads to excced their allocated CU's limit. as I attended one seesion previous week with Micrsoft employee, In that session he explained for example if your capacity is an F8, and your workload takes 100 seconds to complete, and it uses 1500 CU, the scale factor would be calculated as follows: 1500 / 100 / 8 = 1.875. Pls see below image for SKU guardrils for burstable capacity. Read more detail hereBurstable Capacity - Microsoft Fabric | Microsoft Learn 

While coming to your last question, no you ahve to purchase a new capacity and you can't downgrade your reserved capacity until it completes one year. Read more details here Save costs with Microsoft Fabric Capacity reservations - Microsoft Cost Management | Microsoft Learn

Thank you!

Did I answer your question? Mark my post as a solution!

Proud to be a Super User!

[tan_thiamhuat]

notebook executon was around 10to20% high in CU's consumation as compared to dat flowgen2 for same workloads ?? is it high or low?

[suparnababu8 · Accepted solution]

Hi @tan_thiamhuat 

It means that, 10to20% more efficent in CU's consumption. 

Here if you see notebooks consumption of CU's is low as compared to Dataflow gen2. read more about it here Microsoft Fabric – Comparing Dataflow Gen2 vs Notebook on Costs and usability - FourMoo | Microsoft ...

Thank you!

Did I answer your question? Mark my post as a solution!

Proud to be a Super User!

[tan_thiamhuat]

yes, then it should be low in consumption. In the first statement, you mentioned "high" which I think is some typo error.

[suparnababu8 · Accepted solution]

Hi @tan_thiamhuat 

my bad, yes it's typo. Thanks for cathing that😎

Did I answer your question? Mark my post as a solution!

Proud to be a Super User!

[burakkaragoz · Accepted solution]

Hi @tan_thiamhuat ,

Great set of questions — here's a breakdown based on what we know so far:

1. How are Capacity Units (CU) computed?

CU usage is based on the compute resources consumed during execution. This includes CPU, memory, and I/O operations. The more intensive your operation (e.g., large data transformations, joins, or writes), the more CU it consumes.

2. Notebook vs Dataflow Gen2 CU usage

In general:

• Notebooks tend to be more efficient in CU usage because they give you more control over execution (e.g., batching, filtering early, caching).
• Dataflow Gen2 is low-code and easier to use, but it may abstract away optimizations, leading to slightly higher CU usage for the same task.

There’s no official percentage published, but in practice, Notebooks can use 20–40% less CU depending on how optimized your code is.

3. Bursting and Smoothing

• Bursting allows workloads to temporarily exceed the provisioned CU limit by borrowing from unused capacity (if available). This helps avoid throttling during short spikes.
• Smoothing spreads out CU usage over time to avoid sudden spikes. It’s useful for workloads that have uneven demand patterns.

These mechanisms help maintain performance without over-provisioning.

4. Changing reserved capacity

If you purchase a 1-year reservation with a 40% discount:

• You cannot downgrade the capacity during the term.
• You also cannot upgrade mid-term. You’d need to purchase additional capacity separately.
• Microsoft recommends sizing carefully before committing to a reservation.

If flexibility is important, consider using pay-as-you-go until you’re confident about your workload patterns.

Hope this helps clarify things.

If my response resolved your query, kindly mark it as the Accepted Solution to assist others. Additionally, I would be grateful for a 'Kudos' if you found my response helpful.