Question about F8 Fabric licencing

[PBIjk]

Hi,

I'm struggling a little bit to undrestand how exactly the Fabric capacity licencing works.

My organization is currently using F8 capacity. According to this resource, this give us:

SKU	Capacity Units (CU)	30-second CU use
F8	8	240

A. This, as I understand, means that we have access to 8 CU = 'compute resources'. We can then run our dataflows but the regular speed is whataever can be computed by 8 CU.

This can be extrapolated to mean that we have 240 CU 'horsepower' per 30 seconds, 480 CU per minute, 28.800 CU per hour and 691.200 CU per 24 hours.

B. The A is not the whole story though. As previously assumed, there is around 691.200 CU per 24 hours. If I have a Dataflow that requires 28.800 CU to be completed (which would take 28.800 / (F)8 / 60 = 60 minutes to be completed), the Fabric can 'accelerate' my F8 to e.g. F64, which in result would finish my Dataflow 8x faster (60 / 8 = 7,5 minutes). This is what is called bursting here.

C. Bursting is not a free running wild process though; it's still governed by the Smoothing. This can distribute the workload across time to not hit the Capacity Units limit.

Here is when it get's confusing to me:

1. If I run one big job in capacity that has no other jobs planned for next 24 hours - will it run at max speed (and what is the max speed? F256, F2048?)
   
   

2. In below graph:

What does 100% CU limit line mean? Available resources for 24 hours? (691.200 CU for F8?). Or maybe 8 CU limits currently used by Backrgound jobs in blue?

3. How exactly can one hit capacity limit?

4. The image above (May 6 part) is a representation of the heavy job that was manually ran evening of May 5. It crippled our reporting, as the Report User can't use the reports.

What exactly is happening here? Why wasn't the heavy job not Smoothed out through the remianing 24 hours? No other operations were running since the start of heavy job (apart from users trying to use reports).

5. Are there any monthly limits on CUs?

Thanks!

[v-karpurapud]

Hi @PBIjk 

May I ask if you have resolved this issue? If so, please mark the helpful reply and accept it as the solution. This will be helpful for other community members who have similar problems to solve it faster.

Thank you.

[v-karpurapud · Accepted solution]

Hi @PBIjk 

Thank you for reaching out to the Microsoft Fabric Community Forum. Also, thank you @nilendraFabric for sharing  insights on this topic.

Your understanding of Fabric F8 capacity is mostly correct, but here are some key clarifications:

1. What does the 100% CU Limit line mean?

The 100% CU line in your Fabric utilization chart shows the maximum allowed smoothed compute utilization over short time windows, not a daily consumption cap. Specifically for F8:

10-minute smoothed limit = 4,800 CU (8 CU/sec × 600 seconds)

60-minute smoothed limit = 28,800 CU (8 CU/sec × 3,600 seconds)

This line helps monitor whether you're hitting these sustained throughput limits. It does not correspond to a daily budget like 691,200 CU that figure is theoretical and unthrottled.Even if you're under 691,200 CU in a day, you can still get throttled if your usage exceeds the allowed rate in a short time window.

2. What is the max burst speed?

Fabric can burst beyond your base F8 (8 CU) temporarily, but the actual burst ceiling is not publicly documented. It depends on available capacity in the region, internal service limits, and Microsoft’s dynamic resource allocation

While you might see performance similar to F64 or higher, you cannot rely on a specific burst level.  Bursting only applies to background jobs (like dataflows or notebooks), not interactive workloads like report viewing.

3. How can you hit the capacity limit?

You hit the limit when smoothed CU usage exceeds the 10- or 60-minute thresholds. This can happen even if your total 24-hour usage is low, especially if a large job runs in a short burst and smoothing can’t spread it out enough.

4. Why wasn’t the heavy job smoothed out?

Smoothing in Fabric tries to distribute background job load over 24 hours. However, if a job is too large or completes too quickly, it can still exceed short-term CU thresholds (e.g., 4,800 CU over 10 minutes for F8). This triggers throttling, which deprioritizes interactive operations like report usage even if no other jobs are running.

In your case, the job likely spiked CU usage in a short window, and smoothing couldn’t spread it out fast enough. That’s why reports became slow or inaccessible.

5. Are there monthly CU limits?

No, there are no monthly CU usage limits. However, your Fabric subscription has a quota on the total provisioned capacity units. For example, provisioning an F8 (8 CU) and an F16 (16 CU) would count as 24 CU toward your tenant’s quota (e.g., 192 CU total).

This quota affects how much capacity you can provision, not how many compute units you can use over time. It's an administrative limit, not a performance limit.

If this response resolves your query, kindly mark it as Accepted Solution to help other community members. A Kudos is also appreciated if you found the response helpful.

Thank You!

[PBIjk]

Hi @v-karpurapud ,

Thank you very much for your answer - this make much more sense now. I have some follow-up questions though 🙂 

1. My understanding of smoothing has been (based mostly on this video from KratosBI) that whatever I throw at Fabric that would 'fit in' remaining 'usable' capacity for next 24 hours is fair game. 

However, you mentioned here:

---

Even if you're under 691,200 CU in a day, you can still get throttled if your usage exceeds the allowed rate in a short time window.

And here:

4. Why wasn’t the heavy job smoothed out?

Smoothing in Fabric tries to distribute background job load over 24 hours. However, if a job is too large or completes too quickly, it can still exceed short-term CU thresholds (e.g., 4,800 CU over 10 minutes for F8). This triggers throttling, which deprioritizes interactive operations like report usage even if no other jobs are running.

In your case, the job likely spiked CU usage in a short window, and smoothing couldn’t spread it out fast enough. That’s why reports became slow or inaccessible.

Could you expand on what makes a job too large or complete too quickly in the context of F8 capacity?

2. Suppose I have a 600,000 CU job that I needs to run every other day and an entirely “fresh” F8 capacity. The theoretical un‑throttled 24‑hour window for F8 is 691,200 CU.

Is it possible that the same 600,000 CU job is sometimes smoothed successfully over 24 hours (≈ 208 CU every 30 seconds, safely below the published F8 limit of 240 CU per 30 seconds) but, at other times, is not smoothed as well and casues throttling?

3. I recall that Fabric’s pay‑as‑you‑go pricing is billed per second, yet Microsoft also lists a monthly figure of  US $1,051.20 for F8. Could you clarify how the billing actually works? 

Thanks!

[PBIjk]

@v-karpurapud 

Would really appriciate your feedback on this, thanks 🙂 

[nilendraFabric]

Hi @PBIjk 

grest question and details first of all 

let me give my 2 cents here

Smoothing distributes CU usage over future timepoints:
• Interactive operations (e.g., report interactions) are smoothed over 5–64 minutes.
• Background jobs (e.g., dataflows) are smoothed over 24 hours.
• Throttling triggers when a capacity’s smoothed CU usage exceeds its 10-minute or 60-minute limits. For F8:
• 10-minute limit: 4,800 CU (8 CU/sec × 600 seconds).
• 60-minute limit: 28,800 CU (8 CU/sec × 3,600 seconds).
In your case, the heavy job likely exceeded these limits despite smoothing, causing throttling that blocked new interactive operations (e.g., report access)

There are no monthly CU consumption limits. However, Fabric quotas cap the total CU provisioned across capacities in a subscription (e.g., an F8 + F16 might hit a quota of 192 CU)