What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Measure electricity across a clearly stated system boundary, integrate power over the workload, and divide by a fixed unit of useful work. For a training run, publish total kilowatt-hours (kWh) and identify whether the total includes accelerators, host CPUs and memory, storage, networking, idle capacity and facility overhead. For inference, report watt-hours (Wh) per request or per token volume together with model, hardware, input and output sizes, batching, concurrency and utilization.
A graphics processor’s thermal design power (TDP) multiplied by elapsed time is only a partial estimate. Actual telemetry or a suitable electricity meter is preferable, and every estimate should identify omitted components, geography, time period and uncertainty.
As an Amazon Associate I earn from qualifying purchases.
The measurement in one sentence
Define what is being measured, choose the functional unit before running the workload, collect power or energy readings for the complete interval, integrate those readings, and report both the raw total and the normalized result. A number without its boundary and workload description cannot answer whether one model, run or service is more energy-efficient than another.
Start with a system boundary
The boundary is the equipment and facility activity counted in the result. Write it down before collecting data; changing it after the run can make two apparently precise figures incomparable.
#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
Device, node, cluster or facility
- Device boundary: energy at a single accelerator, desktop or server component. This is useful for component tuning but does not represent the electricity drawn by the complete system.
- Node boundary: one server, including its accelerators, CPUs, memory, fans, motherboard and local storage. It is often practical for a workstation or dedicated server.
- Cluster boundary: all nodes assigned to the job, plus the network and shared storage that the job uses. Include scheduler and management equipment when they are inside the declared allocation.
- Data-center boundary: information-technology energy plus the facility overhead represented by power usage effectiveness (PUE). This is the appropriate scope when making an operational electricity or carbon claim about a hosted service.
Also state whether the result is operational electricity only or a broader life-cycle assessment. ITU-T L.1801’s standards-oriented framework calls for an explicit goal, scope and measurable functional unit, with data collected across the relevant life-cycle stages when a life-cycle study is intended.
Choose the functional unit first
| Workload | Recommended unit | Required context |
|---|---|---|
| Completed training run | Total kWh per run | Model and task, dataset or token count, hardware and count, precision, completion criteria, duration, and included components |
| Single inference request | Wh per request | Model version, prompt and output token counts, response quality target, batch or concurrency setting, hardware and idle allocation |
| Service throughput | Wh per 1,000 tokens | Whether tokens are input, output or combined; tokenizer and accounting convention; traffic mix, batching, concurrency and utilization |
| Carbon result | kg CO2e per run or request | Measured energy, PUE boundary, emissions-factor source, geography and time resolution |
Keep the quality target and workload constant when comparing systems. A shorter response, smaller prompt or lower-quality model can reduce energy without demonstrating a more efficient implementation.
Collect actual energy data
Telemetry is better than a nameplate rating
Use integrated energy counters or high-frequency power telemetry from the accelerator, server, rack or provider whenever possible. Sum each reading over its interval:
Energy (kWh) = Σ [power (kW) × interval (hours)]
If readings are in watts and seconds, divide the product by 3,600,000 to obtain kWh. Record timestamps so that the measurement starts before initialization and ends after the final output, checkpoint or shutdown activity included in your run definition.
TDP describes a thermal design target, not the electricity consumed at every moment. Multiplying TDP by runtime omits host power, memory, fans, power-supply losses, storage, networking and facility overhead. It can provide a rough upper or lower bound only when the omission and interpretation are clearly labeled.
Use a wall meter for a suitable single system
A plug-in electricity usage monitor can measure whole-system wall energy for a compatible workstation or test bench. Check its electrical rating and outlet compatibility before connecting it. The reading includes everything on that outlet, so it cannot isolate one GPU, attribute energy among concurrent jobs or add data-center overhead. It is therefore useful for a dedicated local test, not a cluster-wide production accounting method.
Use provider and cluster telemetry carefully
For cloud or multi-node jobs, collect accelerator and host readings from the provider or orchestration layer and identify which values are measured and which are modeled. Ask whether the data includes idle machines reserved for the service, CPU and RAM, network and storage, and facility overhead. If the provider supplies only active accelerator power, report that narrower boundary rather than calling it total service energy.
Free tools Windows power users keep installed
One-click scans. No signup required.
Measure a training run
Training is usually bounded by a start event and a completion event, but the definition must cover work that materially contributes to the result.
- Define completion. Record the model, training task, dataset or token target, precision, target quality metric and the checkpoint that counts as the completed run.
- List hardware. Record accelerator type and count, host CPU and memory configuration, storage, network links and software or runtime versions that affect utilization.
- Synchronize clocks. Ensure telemetry timestamps and job logs use the same time basis. Capture initialization, data loading, evaluation, checkpointing and final shutdown if those activities are inside your run boundary.
- Integrate readings. Sum energy for every node or meter over the defined interval. Preserve the raw per-interval data so another analyst can audit the total.
- Account for shared equipment. If storage, network or management systems serve multiple jobs, use an allocation rule and describe it. If no defensible allocation is available, mark those components as excluded rather than silently assigning zero.
- Handle interruptions consistently. Decide whether failed, pre-empted or restarted attempts count. A result for the energy needed to obtain the final checkpoint should include discarded attempts that were necessary to reach it; a result for one successful attempt can report them separately.
- Publish raw and normalized values. Give total kWh for the run, then any secondary intensity such as kWh per training token or per completed example, with the denominator and quality target.
What to do with failed and restarted jobs
Keep a run ledger containing start and stop times, reason for termination, checkpoint status and measured energy. Report the final successful attempt and the cumulative energy to obtain the published model as separate quantities when failures or restarts occurred. This prevents a low figure for one clean attempt from being mistaken for the energy actually spent in production.
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
Estimating when direct measurement is unavailable
If you have only accelerator power and runtime, calculate that component and label it an estimate. State that host, storage, network, idle and facility energy are omitted or separately modeled. Do not convert the partial result into a whole-system claim.
Measure inference as a service workload
Inference is repetitive and its energy depends strongly on request shape and serving policy. A request-level result is meaningful only when those variables are recorded.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsWh per request
For each test period, measure the service boundary over many requests, then divide total energy by the number of completed requests:
Wh/request = total measured Wh ÷ completed requests
Report median or average only with the distribution or sample description. Include prompt-token and output-token counts, model and precision, accelerator type, batch size, concurrency, response length limits and quality settings. A short prompt and short answer should not be presented as representative of a longer production conversation.
Wh per 1,000 tokens
Token normalization can aid comparison, but specify whether the denominator is input tokens, output tokens or their sum, and name the tokenizer or counting convention. State whether requests were batched and whether the service was operating at a fixed concurrency or a production traffic mix. Without these details, two “Wh per 1,000 tokens” figures may measure different work.
Allocate idle serving capacity
A production endpoint can consume electricity while waiting for requests. Decide whether the boundary covers active accelerator work only or the full service capacity reserved for customers. For a production boundary, allocate idle machines and baseline CPU, RAM and cooling energy across the requests or tokens served during the interval. State the allocation rule, especially when utilization changes throughout the day.
Account for every component in the declared boundary
| Component | Measurement question | Common omission |
|---|---|---|
| Accelerators | Are readings actual board or device energy, and are all devices included? | Counting only the busiest device or using TDP as if it were measured energy |
| Host CPU and memory | Is server power measured at the node or modeled from utilization? | Reporting accelerator energy as system energy |
| Storage and network | Are local disks, shared storage and fabric switches inside the boundary? | Ignoring data movement and checkpoint traffic |
| Idle capacity | Are reserved but underused machines allocated to requests or runs? | Dividing active energy by throughput while leaving standby power unassigned |
| Facility overhead | Is PUE applied, and is its source and period known? | Comparing a wall or device figure with a data-center figure |
Google’s 2025 inference methodology illustrates a comprehensive service boundary: it includes dynamic system power, idle machines, CPU and RAM, and data-center overhead. A narrower active-TPU/GPU-only calculation gives a different answer even for the same service.
Worked calculations
Integrating interval power
Suppose a node is measured at 4.0 kW for 30 minutes and 3.0 kW for the next 90 minutes. The integrated energy is (4.0 × 0.5) + (3.0 × 1.5) = 6.5 kWh. If a second identical node runs over the same interval, the cluster total is 13.0 kWh before any facility overhead. These numbers illustrate the arithmetic; your report should use the readings from the actual run.
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.
Normalizing inference energy
If a defined service boundary consumes 2,400 Wh while completing 12,000 requests, the measured intensity is 0.2 Wh per request. If those requests contain 6,000,000 counted tokens in total, the same interval is 0.4 Wh per 1,000 tokens. The two units answer different questions and should be published with the request mix and token convention.
Applying PUE
When IT equipment energy is 100 kWh and the declared facility PUE is 1.20 for the relevant location and period, operational facility energy is 120 kWh. Identify the PUE source and whether it is a measured annual value, a provider estimate or another basis. Do not apply a generic PUE to a device-only result without saying that the facility figure is modeled.
Keep carbon separate from energy
Energy is measured in kWh; carbon is calculated using an emissions factor. A practical operational formula is:
CO2e = IT energy (kWh) × PUE × grid emissions factor (kg CO2e/kWh)
The emissions factor must match the geography and time period of the electricity. Specify whether it is location-based or market-based, which dataset supplied it, and its time resolution. A request served in a different region or hour can have the same measured Wh but a different carbon result. Include uncertainty in both the energy and emissions-factor inputs.
Crashes, 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 minuteWindows 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 reinstallQuantify uncertainty instead of implying false precision
Separate measured values from estimates in the report. Identify sensitivity to:
- accelerator and host telemetry accuracy;
- the start and end timestamps;
- utilization and batching;
- allocation of idle capacity;
- shared storage and network energy;
- PUE and its measurement period; and
- the geographic and temporal emissions factor.
Run a sensitivity calculation when an assumption could change the conclusion. For example, present a low and high result using alternative idle-allocation or PUE assumptions, and explain which components drive the range. If only a modeled estimate is possible, call it an estimate and avoid extra decimal places that the inputs cannot support.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Make comparisons apples-to-apples
Before ranking models, hardware or serving configurations, align every item below:
- functional unit and denominator;
- task, quality target and completion criteria;
- model version, precision and software runtime;
- accelerator type, count and host configuration;
- duration, utilization, batch size and concurrency;
- prompt, output and dataset characteristics;
- treatment of host, idle, storage, network and facility energy;
- PUE, geography and time period; and
- measured versus estimated method and uncertainty.
If one figure uses active accelerator energy and another uses full production energy, describe the boundary difference rather than presenting a winner. Likewise, a per-request value from a low-concurrency test is not directly comparable with a heavily batched throughput benchmark.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →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
Published figures and how to interpret them
Google published two 2025 estimates for a median Gemini Apps text prompt. They demonstrate why methodology matters, not a universal energy requirement for AI.
| Google figure | Boundary and qualification | What it shows |
|---|---|---|
| 0.24 Wh per median text prompt | Google’s comprehensive methodology; includes dynamic system power, idle machines, CPU and RAM, and data-center overhead | A fuller service estimate for that product, workload and methodology |
| 0.10 Wh per median text prompt | Google’s active TPU/GPU-only comparison; omits system elements included in the comprehensive method | How a narrower boundary lowers the reported number |
Do not compare either value with another provider’s prompt figure unless model behavior, input and output sizes, hardware, utilization, idle allocation and system boundary match. Google Cloud’s August 21, 2025 explanation quotes Amin Vahdat and Jeff Dean: “Many current AI energy consumption calculations only include active machine consumption, overlooking several of the critical factors discussed above.”
A practical reporting template
Use the following fields for every published result:
- Question and scope: state whether the result covers a device, node, cluster or data center, and whether it is operational electricity or life-cycle energy.
- Functional unit: give kWh per completed training run, Wh per request or Wh per 1,000 tokens, and define the denominator.
- Workload: identify model, task, quality target, dataset or token counts, prompt and output sizes, batching and concurrency.
- Hardware and software: list accelerator and host types, counts, precision, runtime and relevant configuration.
- Measurement: name the meter or telemetry source, sampling interval, integration method, start and stop events, and whether readings are measured or modeled.
- Included components: explain treatment of accelerator, CPU, RAM, storage, network, idle capacity and facility overhead.
- Results: publish raw total energy, normalized intensity, sample size and any distribution or uncertainty range.
- Carbon, if provided: list PUE, emissions-factor source, geography, time period and calculation basis.
Troubleshooting common measurement failures
The number is implausibly low
Check whether only accelerator power was counted, whether host and memory telemetry was excluded, and whether idle machines or PUE were omitted. Verify that initialization, checkpointing and evaluation were not cut from the timestamps.
Two runs with the same model disagree
Compare prompt or dataset size, output length, batch and concurrency, precision, clock limits, thermal throttling and utilization. Confirm that both runs use the same boundary and that one did not include standby capacity or facility overhead.
Cloud telemetry has gaps
Mark missing intervals, obtain provider-level energy data if available, or model the gap with a documented assumption. Do not silently interpolate and present the result as fully measured.
A wall meter conflicts with device telemetry
The wall meter includes power-supply losses and all equipment on the outlet; device telemetry may cover only one board. Check sampling windows, host load and other processes, then report the two boundaries separately rather than forcing them to match.
Standards and methods to consult
ITU-T L.1801 provides a framework for defining goals, scope and functional units and for gathering life-cycle data. ITU’s 2025 discussion of AI energy measurement identifies actual measurement as best practice and treats real-time telemetry as more accurate than relying on TDP. The Green Software Foundation’s SCI for AI methodology emphasizes publishing intensity alongside absolute totals. IEEE P7100 is an active standards project concerned with AI energy and environmental considerations; verify its current status and applicable draft before citing a finalized requirement.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →These frameworks help structure a measurement, but they do not create one universal number. Your result remains specific to the stated workload, boundary, location, time period and method.
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.
