In March 2025, Tailpipe updated its methodology for calculating the power draw of central processing units (CPUs). Previously, Tailpipe based CPU power consumption on third-party testing. To increase accuracy, Tailpipe ran its own measurement of 37 CPU configurations to obtain specific, verifiable, and up-to-date data.
To calculate the power draw of CPUs, Tailpipe first matches the customer’s virtual machine name to the Tailpipe database of virtual machine specifications. Tailpipe has collated a central database that details every instance hosted by in-scope cloud service providers (CSPs), along with their associated variable components (CPU, RAM, storage, and GPU). This database was primarily populated by querying CSPs’ Command Line Interfaces, scouring their online documentation and product specifications, with third-party databases from Boavizta and Teads Engineering providing secondary data in the rare cases where CSP data was unavailable. See Tailpipe’s Data Quality Dashboard for more information on our data sources.
The database details the model and specification of the CPU used by the cloud provider to run each instance. A key aspect of cloud computing is its capability to share infrastructure with other customers, reducing costs and improving efficiency. That infrastructure resource sharing includes the CPU. Because the total computing power of a single CPU within the overall computing platform is usually greater than that of a single virtualized instance used by a single customer, platform CPUs are often shared by multiple instances across multiple customers.
Another way this is described is by referring to virtual CPUs (vCPUs). One vCPU represents a single thread on a CPU. Customers purchase virtualized instances with quantities of vCPUs that can carry out the customers’ computing workloads. To calculate what percentage of a single CPU is being used by an instance – the resource share, as it’s called in the Software Carbon Intensity Specification – Tailpipe divides the number of vCPUs the instance has by the number of CPU threads that the platform machine hosting the instance has.
For example, an AWS c5a.8xlarge instance has 32 vCPUs. The c5a.8xlarge is hosted on a c5a.24xlarge machine with an AMD EPYC 7R32 CPU, which has 96 threads. The c5a.8xlarge therefore uses 32/96 – or 1/3 – of the platform CPU’s computing power.
To calculate the power draw of the CPU, Tailpipe first references the amount of power (in Watts) that a CPU consumes at different levels of utilization. Tailpipe measures the power consumption of CPUs using a calculation based on Thermal Design Power (TDP). TDP is the maximum amount of heat (measured in Watts) that a CPU can produce at 100% utilization and is published by chip manufacturers. TDP is therefore a reasonable approximation of the total power required by the CPU. For example, the AMD EPYC 7R32 CPU has a TDP of 280 Watts. This means that the 7R32 CPU will use no more than 280 W at peak utilization. As utilization decreases, power consumption decreases in a pattern consistent across CPUs.
Tailpipe ran stress tests on a range of CPUs within bare metal instances running Ubuntu Linux in cloud service providers (bare metal instances are those where the customer has access to all the resources within the platform). Tailpipe varied the utilization of the CPUs using stress-ng and measured their precise power consumption using Turbostat. In this way, Tailpipe took measurements of the power consumption in Watts of the CPUs at 5% utilization increments.
By analyzing the results from 37 discrete bare metal platforms, Tailpipe found the following relationship between CPU Utilization and Power Consumed by the CPU expressed as the Percentage of TDP:
| CPU Utilization (%) | TDP Power Factor (%) |
|---|---|
| 0 | 24.4 |
| 5 | 32.0 |
| 10 | 39.2 |
| 15 | 45.9 |
| 20 | 52.3 |
| 25 | 58.2 |
| 30 | 63.7 |
| 35 | 68.7 |
| 40 | 73.4 |
| 45 | 77.6 |
| 50 | 81.5 |
| 55 | 84.9 |
| 60 | 87.8 |
| 65 | 90.4 |
| 70 | 92.5 |
| 75 | 94.2 |
| 80 | 95.6 |
| 85 | 96.4 |
| 90 | 96.9 |
| 95 | 96.9 |
| 100 | 96.6 |
Power consumption of a CPU at a specific level of utilization can then be determined by multiplying the relevant TDP Power Factor by the published TDP of the CPU.
For example, to calculate the power consumption of an AMD EPYC 7R32 CPU with a TDP of 280 that is being 75% utilized:
- Tailpipe finds the TDP power factor at 75% CPU utilization (94.2%, or 0.942)
- It multiplies this 0.942 power factor by the CPU’s TDP: 0.942 * 280 = 263.76
This CPU therefore consumes 263.76 W at 75% utilization.
In an operational environment, Tailpipe retrieves the actual CPU utilization data from the customer’s account (for example AWS CloudWatch or Azure Monitor). This details the CPU utilization for every hour of the usage period. Tailpipe then matches the CPU utilization value for each hour to the wattage drawn at that specific level of utilization.
Tailpipe totals each CPU power consumption Watt value for each hour, to deliver the entire energy consumption of the CPU over the measured time period. Finally, Tailpipe factors in the ratio of the instance vCPU as a proportion of the CPU’s total computing power. The calculation for CPU power draw is therefore:
(TDP * CPU Utilization TDP Power Factor) * (instance vCPU/CPU threads)
Using the c5a.8xlarge as an example, the CPU power draw of this instance at 75% utilization would be:
(280*0.942) * (32/96) = 87.92 Wh
If this instance was being utilized at 75% for a full day, it would consume 2,110.08 W.
