iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Measure energy across a declared system boundary, integrate actual power over the workload, and divide by a clearly defined unit of work. For a training run, report total kilowatt-hours (kWh) for the complete run. For inference, report watt-hours (Wh) per request or per a stated token quantity, together with model, hardware, prompt and output sizes, utilization, and batching. A GPU’s thermal design power (TDP) multiplied by runtime is only a partial estimate, not measured whole-system energy.
The number is meaningful only when a reader can tell what was measured, where, when, and under which workload conditions. The same model can produce very different results if one measurement includes host CPUs, memory, idle serving capacity, networking, storage, and facility overhead while another counts only active accelerators.
Start by defining exactly what you are measuring
Write the measurement question before collecting telemetry. “How much energy does an AI model use?” is too broad to produce a reproducible answer. A useful question names the workload, boundary, and functional unit, such as “How many kWh did this completed fine-tuning run consume at the node boundary?” or “How many Wh did this production inference service use per request, including allocated idle capacity and data-center overhead?”
Choose the system boundary
State whether the boundary is a single device, a server node, a multi-node cluster, a cloud instance, or an entire data center. Also state whether you are measuring operational electricity only or attempting a broader life-cycle assessment. ITU-T L.1801’s standards-oriented framework calls for an explicit goal, scope, functional unit, and collection of data across relevant life-cycle stages when those stages are in scope.
#1 Best Overall
- Various Monitoring Parameters: The power meter plug can monitor the power (W), energy (kWh), volts, amps, hertz, power factor, cost, minimum and maximum power (W), cumulative days and time of your appliances. By switching 7 display modes, you can easily know the various parameters while the appliance is working. The home energy monitor can also calculate and display how much power your appliance uses and how much electricity bill it cost in cumulative time
- Upgraded LCD Display: With large screen size 2.36 inch x 1.85 inch, clearer monitor backlit, our electrical usage monitor can display the data clearer and more visible no matter day or night. 180°full wide viewing angles is great for reading and recording the data in any angles. No need to stand on the front of the display and bend over to read the numbers
- Adjustable Backlight Time: Our upgraded watt meter has 5 options of backlight time. The default backlight time duration is 10 minutes(bL-0). If you want to change the backlight time, you can press and hold "UP" and "DOWN" button at the same time to enter backlight time setting, then press "UP" and "DOWN" to select the backlight time (bL-0 =10 minutes, bL-1=1 hour, bL-2=4 hours, bL-3=8 hours, bL-4=always on), finally press the "COST" to save the backlight time settings
- Overload protection: When the power of the appliance exceeds the overload power, the LCD will display “OVERLOAD” to warn the user. All the buttons will quit working and can only be workable when you lower or remove the load power. The default overload power is 3680W and is adjustable from 0 to 3680W. In general, you need to set the overload power to 1800W before using. Just press the "function" button for more than 3 seconds to enter the setting
- Data Memory Function: The wattage meter will record your power consumption data when you remove it from socket, or remove appliances from the electricity monitor. You can directly see the last data when you use it next time. This function can also automatically save the data when there is a sudden power failure
A practical operational boundary can be described in layers:
- Accelerator only: GPU, TPU, NPU, or another AI accelerator. This is useful for hardware studies but is not a production-service total.
- Node or server: accelerators plus host CPUs, memory, motherboard, fans, power supplies, and local storage.
- Cluster or allocated service: all participating nodes, shared storage and networking, and any capacity reserved for the workload.
- Facility: the IT load plus data-center overhead represented by power usage effectiveness (PUE).
Do not silently switch boundaries between experiments. If a value covers active accelerator power only, label it that way in the result rather than calling it “the energy used by the model.”
Pick a functional unit before taking readings
The functional unit is the denominator that makes an energy result interpretable.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute| Workload | Recommended primary unit | Details that must accompany it |
|---|---|---|
| Completed training run | kWh per run | Model and task, data set or token count, hardware and count, completion criteria, duration, software conditions, and boundary |
| One inference request | Wh per request | Model, prompt/input length, output length, batching, concurrency, response quality or decoding settings, utilization, and boundary |
| Token-normalized inference | Wh per 1,000 tokens | Whether tokens mean input, output, or input plus output; tokenizer; workload mix; batching; and service boundary |
| Carbon result | kg CO₂e per run or request | Energy total, PUE, emissions-factor source, geography, time period, and calculation boundary |
Wh per request is easy to understand but changes with prompt and response length. Wh per 1,000 tokens can improve comparisons only when the token convention and workload are identical. A short prompt with a short answer is not comparable to a long-context, long-output request simply because both are called “one inference.”
Measure energy rather than inferring it from TDP
Actual energy telemetry or an electricity meter is preferable. If you have power readings, integrate them over the measurement interval: energy (kWh) = Σ [power (kW) × interval (hours)]. With one-minute samples, convert each watt reading to kilowatts and multiply by 1/60 hour, then sum every interval from the defined start to the defined end.
TDP is a thermal and design specification, not a record of the power drawn by a particular workload. Multiplying accelerator TDP by runtime ignores utilization changes and omits host CPUs, memory, fans, power-conversion losses, storage, networking, idle machines, and facility overhead. ITU’s 2025 reporting describes real-time measurement as best practice and regards TDP-based calculations as less accurate than actual telemetry.
Use a meter for a suitable single-system test
A plug-in electricity-usage monitor can measure wall energy for a suitable single workstation or test bench. Verify its electrical rating, voltage, current capacity, and outlet compatibility before connecting equipment. Put the monitor between the outlet and the complete system, record the starting reading, run the workload, and record the ending reading. The difference is whole-system wall energy for that test boundary.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →This method cannot isolate GPU energy, attribute consumption among concurrent jobs, or include data-center overhead. It is therefore appropriate for a dedicated local workstation, not a shared cluster or cloud service unless the provider supplies equivalent boundary data.
Rank #2
- Multi-function power monitor: Our electric usage monitor can monitor the power (W), electricity(kWh), voltage(V), frequency(Hz), current(A), power factor(PF), unit price($/kWh), total cost($) of your appliances. By switching 8 display modes, you can easily know the various parameters while the appliance is working. The “electricity” mode can calculate and display how much power your appliance uses. And the “total cost” mode will show how much electricity bill it cost in cumulative time
- Overload protection: When the loading power of the appliance is over the default overload threshold 1800W, the whole display with backlight and the word “OVERLOAD” will keep flashing to warn the users. Please turn off the appliance for safety concern
- Premium Material: The whole body of our energy meter is made of high-quality ABS material. It makes our home electricity usage monitor more long lasting, fireproof and anti-drop. The standard US socket and plug is suitable for all US standard appliances
- Backlight Display: With white backlight and black words, the LCD display of our home power monitor can display the data clearer and help you to read the data easier no matter day or night. The backlight will only lights up when the device is connected to AC power. If no button is pressed, the backlight turns off automatically after 10 minutes. You can also press the "UP" button to turn off the backlight manually, and press any button to turn on backlight again
- Easy to reset: No reset tool needed, our appliance power usage meter is easy to reset. You can press the “M” button for 5 seconds directly to reset the device. After reset, all cumulative data (electricity quantity, cost) will be cleared, and all settings will be restored to factory settings
Use telemetry for nodes and clusters
For servers, collect accelerator and host energy counters when available, and retain timestamps at a consistent sampling interval. Record the software runtime, driver and accelerator configuration, number of devices, and job identifier so readings can be joined to the workload. For a cluster, sum participating nodes and document treatment of shared storage, network equipment, and reserved capacity.
Where a provider exposes only power rather than cumulative energy, integrate its samples. Handle gaps explicitly: estimate missing intervals, report the method, and mark the affected component as estimated. Do not present a reconstructed value with more decimal places than the telemetry supports.
Measure a complete training run
Training is normally bounded from job start to successful completion, but the run definition must say what happens to setup, evaluation, checkpointing, failures, and restarts. A reproducible report distinguishes the energy of the successful run from energy spent on abandoned or repeated attempts.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →- Define the completion event. Record the exact start condition and the event that ends the run, such as the final checkpoint after a stated number of steps or epochs. Decide whether data loading, warm-up, validation, checkpoint writing, and final evaluation belong inside the boundary.
- Record the workload. Identify model architecture and parameter size, precision, training objective, data set or token volume, batch size, sequence length, optimizer, number and type of accelerators, host configuration, and software/runtime versions.
- Start energy collection before initialization. Capture the baseline or starting meter value before the process allocates devices. This prevents setup and loading energy from disappearing from the total.
- Run without changing the measurement scope. Keep the same device count, power policy, batch/concurrency settings, and placement assumptions. Log interruptions, throttling, checkpoint intervals, and restarts.
- Stop at the declared completion event. Include the final checkpoint and any explicitly included evaluation. Record the ending meter or cumulative telemetry value.
- Integrate and reconcile. Calculate total kWh, compare it with per-component readings, and investigate large differences. A wall-meter total should normally exceed accelerator-only energy because it includes the rest of the system.
- Publish both total and context. Report kWh per run first, then any normalized value such as kWh per training token or per completed example. Keep the run definition stable when comparing experiments.
Handle failed and restarted jobs
Failures can materially change total energy. You can report the successful run alone, or report “project energy” that includes failed attempts, but name the choice. For the latter, list each attempt’s duration and energy and sum them. Never imply that a successful-run figure includes restarts when it does not.
Measure inference as a repeated service workload
Inference energy depends on request shape and service behavior. Prompt length, generated output, model path, precision, batching, concurrency, queueing, accelerator utilization, and idle capacity all affect Wh per request.
- Freeze the request specification. Define the model version, tokenizer, input and output token accounting, decoding settings, maximum context, and quality target. Use a representative request set rather than a single unusually short prompt.
- Define concurrency and batching. Record simultaneous requests, batch formation, continuous batching behavior, and rate limits. A server at high utilization usually allocates energy differently from a lightly loaded server with the same accelerator.
- Choose the service boundary. Decide whether the result covers active accelerator energy, the full node, or production service energy including idle machines and facility overhead. State how capacity held available for requests is allocated to each request.
- Measure over a long enough window. Collect energy and request logs over many requests or a fixed time period. Capture total input tokens, output tokens, successful requests, errors, and latency during the same interval.
- Normalize only after aggregation. Calculate
Wh/request = total Wh ÷ completed requests. For token intensity, calculateWh/1,000 tokens = total Wh ÷ total tokens × 1,000, then state whether total tokens means input plus output or another convention. - Publish the distribution when useful. The mean can hide long-context or high-output requests. Report workload mix and, where service decisions depend on it, percentile energy by request class.
Allocate idle serving capacity transparently
Production systems consume energy while waiting for work. Google’s comprehensive Gemini Apps methodology includes dynamic system power, idle machines, CPU and RAM, and data-center overhead. An active-accelerator-only method excludes those elements and will produce a lower number. If you allocate idle energy, describe the rule—for example, proportional allocation across requests in the measurement window—and apply it consistently.
Calculate carbon separately from energy
Energy is measured in kWh; carbon is a separate calculation. A simplified operational calculation is:
Free tools Windows power users keep installed
One-click scans. No signup required.
CO₂e = system energy (kWh) × PUE × grid emissions factor (kg CO₂e/kWh)
Rank #3
- 【Matter-Certified】Matter-certified devices, regardless of brand, can work together and are compatible with most major smart home platforms like Amazon Alexa, Apple HomeKit, Google Home, and Samsung SmartThings. Enjoy more flexible and unified control.
- 【Insightful Energy Tracking】 Monitor your energy consumption with in-depth statistics and clear visuals, helping you optimize power usage.
- 【Estimate Your Energy Bill】 Enhance energy management by integrating with billing systems for clear cost visualization (both single and periodic readings). Additionally, programmable scheduling allows automatic operation of high-consumption devices during off-peak hours with lower electricity rates, resulting in cost savings.
- 【Overcharge Prevention & Power Management】 Automatically cuts off power based on user-set thresholds and durations to prevent overcharging, conserve energy, and protect connected devices from overcurrents by shutting off when power exceeds set limits.
- 【Voice & Remote Control】 With built-in support for both Alexa and Google Assistant, issue simple voice commands to adjust settings, turn devices on or off, or even access specific functions without lifting a finger. Manage Tapo P110M and its connected devices from anywhere with the user-friendly Tapo app.
Use an emissions factor appropriate to the location and time period, and name its source and temporal resolution. A location-based factor, a market-based factor, and a renewable-energy claim can produce different results. PUE must match the facility boundary; do not multiply by PUE if facility overhead is already included in the measured wall energy.
Report the energy result even when carbon factors are uncertain. A carbon estimate without geography, date, factor source, or facility treatment cannot be reproduced.
How to compare two measurements fairly
Before ranking systems, align the following axes:
- Functional unit and token-counting convention.
- Task, data set, prompt/output distribution, and quality or completion target.
- Model version, precision, quantization, batch size, and concurrency.
- Accelerator type and count, host CPUs, memory, storage, and network configuration.
- Duration, utilization, throttling, and idle-capacity allocation.
- Inclusion or exclusion of host, storage, network, and facility energy.
- PUE, geography, and measurement time period.
- Actual meter or telemetry data versus estimates, including missing intervals.
If any axis differs, explain the difference rather than presenting an apples-to-apples ranking. A lower accelerator-only figure can still represent higher production energy if it omits idle servers and facility overhead.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Published figures illustrate why boundaries matter
Google reported two 2025 estimates for Gemini text prompts. Its comprehensive estimate was 0.24 Wh per median Gemini Apps text prompt and included full-system dynamic power, idle machines, CPU and RAM, and data-center overhead. Google also reported 0.10 Wh per median Gemini text prompt using an active TPU/GPU-only method that omitted several elements in the comprehensive method.
These are Google’s service- and methodology-specific estimates, not universal values for AI prompts. They should not be compared with another provider’s figure unless model, request distribution, hardware, utilization, boundary, and allocation method match.
Amin Vahdat, Google’s SVP and Chief Technologist for AI and Infrastructure, and Jeff Dean, Chief Scientist at Google DeepMind and Google Research, summarized the issue in Google Cloud’s August 21, 2025 article, “How much energy does Google’s AI use? We did the math”: “Many current AI energy consumption calculations only include active machine consumption, overlooking several of the critical factors discussed above.”
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Report results so another team can reproduce them
A useful measurement record contains the following fields:
| Field | What to record |
|---|---|
| Workload identity | Model, task, version, quality target, data or request set, and completion criteria |
| Hardware | Accelerator and count, host CPU and memory, node count, storage, network, and power-management settings |
| Workload shape | Training steps/tokens or inference requests, input/output tokens, batch size, concurrency, and utilization |
| Measurement | Meter or telemetry source, sampling interval, start/end timestamps, integration method, and missing-data treatment |
| Boundary | Accelerator, node, cluster, service, or facility; included and excluded components; idle allocation |
| Result | Raw total kWh or Wh, normalized value, units, rounding, and uncertainty or range |
| Carbon | PUE, emissions-factor source, geography, date, and whether the calculation is location- or market-based |
Pair intensity with absolute totals. The Software Carbon Intensity (SCI) methodology recommends reporting intensity alongside absolute emissions or energy context so a lower per-request value is not mistaken for lower total impact when usage grows.
Rank #4
- Various Monitoring Parameters: The power energy meter can monitor the power (W), energy (kWh), volts, amps, hertz, power factor, cost, minimum and maximum power (W), cumulative days and time of your appliances. By switching 7 display modes, you can easily know the various parameters while the appliance is working. The home energy monitor can also calculate and display how much power your appliance uses and how much electricity bill it cost in cumulative time
- Upgraded LCD Display: With large screen size 2.36 inch x 1.85 inch, clearer monitor backlit, our electrical usage monitor can display the data clearer and more visible no matter day or night. 180°full wide viewing angles is great for reading and recording the data in any angles. No need to stand on the front of the display and bend over to read the numbers
- Adjustable Backlight Time: Our upgraded watt meter has 5 options of backlight time. The default backlight time duration is 10 minutes(bL-0). If you want to change the backlight time, you can press and hold "UP" and "DOWN" button at the same time to enter backlight time setting, then press "UP" and "DOWN" to select the backlight time (bL-0 =10 minutes, bL-1=1 hour, bL-2=4 hours, bL-3=8 hours, bL-4=always on), finally press the "COST" to save the backlight time settings
- Overload Protection: When the power of the appliance exceeds the overload power, the LCD will display “OVERLOAD” to warn the user. All the buttons will quit working and can only be workable when you lower or remove the load power. The default overload power is 3680W and is adjustable from 0 to 3680W. In general, you need to set the overload power to 1800W before using. Just press the "function" button for more than 3 seconds to enter the setting
- Data Memory Function: The wattage meter will record your power consumption data when you remove it from socket, or remove appliances from the electricity monitor. You can directly see the last data when you use it next time. This function can also automatically save the data when there is a sudden power failure
Understand uncertainty and common failure modes
“GPU power × runtime” is treated as the final answer
Why it fails: It omits non-GPU components and assumes a constant power draw. Fix: use cumulative energy telemetry or wall measurement; if an estimate is unavoidable, label it accelerator-only and list omitted components.
The denominator is vague
Why it fails: “per prompt” can mean radically different token volumes. Fix: define input and output token conventions, request mix, and quality settings.
Idle energy disappears
Why it fails: a production service may keep machines ready between requests. Fix: state whether idle capacity is included and document the allocation rule.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsTraining boundaries move between runs
Why it fails: one result includes evaluation and checkpointing while another stops at the final optimizer step. Fix: use the same start and completion events and list inclusions.
False precision hides estimates
Why it fails: interpolated telemetry, uncertain PUE, and variable grid factors do not support many decimal places. Fix: identify measured and estimated components, test sensitivity to utilization, PUE, idle allocation, and emissions factors, and report a range where appropriate.
Shared hardware is attributed incorrectly
Why it fails: concurrent jobs share host and facility energy. Fix: use workload-linked telemetry where available, define an allocation method, and disclose that shared components are allocated rather than directly isolated.
Standards and methods to keep watching
ITU-T L.1801 provides standards-oriented guidance for goals, scope, functional units, and life-cycle data. ITU’s 2025 report emphasizes actual measurement and the limitations of TDP. IEEE P7100 was listed as an active project in the available standards material; check its current status before relying on it as a finalized standard. The Green Software Foundation’s SCI methodology is a living project document, so record the version or access date used for a report.
No single universal kWh figure applies to all training runs or inference requests. The defensible result is the one whose boundary, functional unit, workload, measurement method, and uncertainty are visible.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
