5. Inputs and Evidence Sources

The methodology draws on the following categories of inputs.

5.1 Cloud Usage Data

To produce a figure for a unit of cloud usage, the methodology requires the usage attributes that identify it, being the cloud provider, the service or SKU, the usage period, and the quantity or duration of use. These attributes are read from the customer's cloud usage data, most commonly a cloud usage report. No cost or spend information is required.

5.2 Measured Server Power Data

Whole-server electricity measurements come from the database of published server measurements that the Standard Performance Evaluation Corporation maintains. Each published measurement records alternating current power at graduated workload levels, from full load down to active idle, measured through a calibrated power analyser under published run rules. It also records the processor model, the chip and thread counts, the memory modules installed, and the storage devices of the system measured. Server vendors submit the measurements under run rules that the Standard Performance Evaluation Corporation sets.

5.3 Instance Specification Data

Per-instance vCPU counts, memory allocations, processor models, threads per core, and accelerator attachments come from each cloud provider's own published specifications. Those are the AWS instance type documentation, the Azure virtual machine sizes documentation, and the Google Cloud machine family documentation.

5.4 Processor Specification Data

Processor thread counts, clock speeds, process nodes, and rated power figures come from the chip vendor's own documentation where the chip vendor publishes them. Where the chip vendor does not, those quantities come from third-party hardware specification databases. Where a cloud provider designs its own processor and publishes core counts without publishing a clock speed, a process node, or a rated power, Greenpixie estimates those quantities and records each of them as an estimate. A clock speed a cloud provider leaves unpublished is taken from the processor whose curve the figure is scaled from, as described in Section 7.1.

The per-core factors described in Section 7.1 come from the published material of the company that designs the core. Where that company publishes nothing comparing two generations of its own core, the factor comes from an independent published comparison of the two.

5.5 Memory Component Data

Greenpixie's memory module power reference holds module power figures adapted from memory module manufacturer documentation, for the module generations and capacities cloud hosts are configured with. Host memory topology, being the module slot count and socket count a processor supports, is maintained as a per-processor reference. Each entry in that reference records its source.

5.6 Drive Power Data

The drive power reference Greenpixie holds is a table of electricity figures by drive type and capacity, for NVMe solid state drives, SATA solid state drives, and mechanical hard disk drives. The reference is adapted from drive manufacturer datasheets for the drive families deployed in server hardware. Every entry records the document it came from. Published research into data centre energy consumption gives per-drive figures at a fleet level, which Greenpixie holds as context for that reference. The 2016 and 2024 United States Data Center Energy Usage Reports published by Lawrence Berkeley National Laboratory are the principal sources of that research.

The storage sold with each instance size, being the drive count, the capacity per drive, and what kind of drive it is, comes from the cloud provider specifications named in Section 5.3.

5.7 Accelerator Specification Data

Accelerator rated power and memory capacity come from the published documentation of the company that designs each part, read for the form factor the part is sold in. Where that company publishes no rated power for a part, Greenpixie estimates one and records the figure as an estimate. The form factor a given instance carries comes from the cloud provider's own instance documentation. Greenpixie measures the proportion of rated power an accelerator draws in service by running inference workloads on graphics processing units. Every accelerator takes that proportion.

5.8 Burstable Instance Allowances

Cloud providers publish a baseline utilisation allowance for burstable and shared-core instance sizes, expressed as a percentage of a vCPU or as a fractional vCPU. These allowances come from the cloud provider's own published tables.

5.9 Cloud Infrastructure Research

Mapping a billing usage type to the instance that supplies it requires research into each cloud provider's billing catalogue and naming conventions, carried out as described in the parent cloud methodology.