Measure cooling and AI productivity as separate, explicitly bounded metrics. Use the ISO/IEC 30134-7:2023 Cooling Efficiency Ratio (CER) for cooling energy performance, report facility Power Usage Effectiveness (PUE) separately, and define your own reproducible tokens-per-energy test for each model and workload. A credible result states exactly what was metered, for how long, under which thermal and software conditions, and what counted as a completed token or task.
There is no universally standardized tokens-per-watt definition in the sources available for this topic. Treat it as a workload benchmark, not a property of a server in isolation. Cooling can change both overhead energy and AI throughput, so a fair comparison records facility energy, IT energy, accelerator behavior, output rate and quality together.
As an Amazon Associate I earn from qualifying purchases.
Start with the metric you actually need
“Efficiency” can mean several different things in an AI facility. Decide which question you are answering before installing meters:
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- How efficiently does the cooling plant remove heat? Use CER at a declared cooling-system boundary.
- How much facility energy supports IT equipment? Use PUE, with facility and IT energy measured over the same period.
- How much useful AI output do we obtain for energy? Measure defined tokens or completed tasks against energy at a stated accelerator, server, IT or facility boundary.
- What are the water and carbon consequences? Add WUE and CUE rather than hiding them inside an invented single score.
These metrics answer different questions. A low PUE does not prove that a model produces more useful tokens per joule, and a high token rate does not prove that the cooling system is efficient.
#1 Best Overall
- 3-Phase Power Meter With Real-Time Data: see energy usage and solar generation vs. consumption with instant readings in kWh, Watts, Amps, Volts, Power Factor and cost ($).
- Connection Options- by WiFi (2.4 GHz) or wired Ethernet. NEED HELP? Reach out to Eyedro's Technical Support Team for assistance with the setup of your Eyedro device.
- High Resolution Watt Meter Data: once connected, Eyedro maintains full 24/7 usage history. Supporting fixed, tiered and time-of-use rate structures, MyEyedro features responsive real-time graphs, providing you with hourly, daily, weekly and monthly consumption and costs, as well as estimates of what your current electrical bill will cost.
- Multiple Voltage Options: 600V Delta, 480V/277V Wye, 208V/120V, 3-Phase High-Leg Delta and more.
- Directional Solar & Net Metering Ready: differentiates power consumed vs. generated, perfect for solar, grid-tie and net metering scenarios. To break out grid consumption from solar generation and use the MyEyedro Net Meter plugin 2 x Eyedro monitors are required.
Define boundaries, time periods and exclusions
Draw a one-line boundary diagram before collecting data. Mark utility feeds, on-site generation, transformers, UPS equipment, distribution losses, chillers, pumps, fans, cooling towers, heat-rejection equipment, rack or row cooling, embedded server fans and the IT devices running the workload. Record every meter location and identify loads that are included or excluded.
For PUE, ISO/IEC 30134-2:2026 defines the numerator as total data-center energy and the denominator as IT-equipment energy over the same continuous 12-month period. Imported energy, on-site generation and power-conversion stages must be treated consistently, and excluded loads must be disclosed. A shorter observation can be useful, but label it as interim PUE or partial PUE where applicable; do not present a week or a test run as an annual standard value.
Use the same boundary when comparing two designs. Moving server fans from the IT side to the cooling side, or omitting a pump from one case, can make an apparent improvement that is only an accounting change.
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 errorsMeasure cooling with CER
ISO/IEC 30134-7:2023 is the dedicated Cooling Efficiency Ratio standard. Follow its measurement, calculation, reporting and interpretation instructions for the chosen category. The public abstract states that the standard excludes cooling systems not powered by electricity, such as heat-driven absorption chillers. If such equipment is present, describe it separately instead of implying that the CER value covers it.
Choose the cooling-system boundary
State whether CER covers the central plant only or also distribution and room-level equipment. A defensible record says whether it includes chillers, condenser-water pumps, chilled-water pumps, dry coolers, cooling towers, in-row units, rear-door heat exchangers, direct-to-chip pumps, controls and heat-rejection equipment. If server fans are treated as IT energy, say so; if they are assigned to a cooling boundary, say that instead.
Rank #2
- 3-Phase Power Meter with Real-Time Data: see energy usage and solar generation vs. consumption with instant readings in kWh, Watts, Amps, Volts, Power Factor and cost ($).
- Connection Options- by WiFi (2.4 GHz) or wired Ethernet. NEED HELP? Reach out to Eyedro's Technical Support Team for assistance with the setup of your Eyedro device.
- Multiple Voltage Options: 600V Delta, 480V/277V Wye, 208V/120V, 3-Phase High-Leg Delta and more.
- High Resolution Watt Meter Data: once connected, Eyedro maintains full 24/7 usage history. Supporting fixed, tiered and time-of-use rate structures, MyEyedro features responsive real-time graphs, providing you with hourly, daily, weekly and monthly consumption and costs, as well as estimates of what your current electrical bill will cost.
- This Electric Meter Features Non-invasive Installation: uses split-core current sensors (200A included, 5A–6000A compatible) with no wiring changes, no special tools.
Install and validate meters
- Use appropriately rated electrical meters or power analyzers on each cooling subsystem and on the aggregate cooling boundary. A facility feeder requires equipment designed for that voltage, phase count and fault environment; a consumer plug-in meter is not a substitute.
- Synchronize timestamps for electrical and thermal measurements. Check meter scaling, polarity, phase balance and communication dropouts before the test.
- Where the selected CER method requires thermal transfer data, measure supply and return-fluid temperatures, flow rate and the fluid’s volumetric heat capacity. These quantities describe heat movement, but they do not by themselves reproduce every CER category or calculation instruction.
- Log ambient temperature and humidity, supply-water temperature, rack-inlet temperature, cooling mode, set points, valve positions and alarms. Note economizer operation, tower approach and any maintenance or bypass condition.
- Run a data-quality check for missing intervals, implausible flow values, sensor drift and simultaneous changes in workload or cooling mode. Preserve the raw interval data as well as the final aggregate.
Consult the full ISO/IEC 30134-7:2023 text for the exact formula, category definitions and reporting template. The public abstract is not sufficient to recreate those details safely.
Report CER with operating context
Give the observation interval, climate or outdoor conditions, IT load, cooling mode and included equipment beside the CER result. A short commissioning measurement is an interim indicator, not interchangeable with a representative annual value. If a facility uses both air and liquid loops, report the boundaries separately or explain how they are combined under the standard’s categories.
Report PUE as a separate infrastructure ratio
PUE is calculated as:
PUE = total data-center energy ÷ IT-equipment energy
Use the same continuous 12-month period for both terms when claiming an annual PUE under ISO/IEC 30134-2:2026. Include power conversion and distribution consistently, and document on-site generation, imported energy and excluded loads. PUE captures all non-IT overhead inside the declared facility boundary; it is not a cooling-only ratio and it says nothing about model quality or useful tokens.
ISO/IEC 30134-2:2026 explicitly states: “The ISO/IEC 30134 series does not specify limits or targets for any KPI and does not describe or imply, unless specifically stated, any form of aggregation of individual KPIs into a combined nor an overall KPI for data centre resource usage effectiveness or efficiency.” Do not turn PUE and CER into a made-up composite score.
Rank #3
- WITH TRANSPARENT SHELL-- This power meter adopts transparent shell, which is convenient to observe the operation status of electronic components inside the meter.
- LCD DISPLAY WITH BACKLIGHT-- This energy meter is equipped with LCD backlight display is used to read data clearly, convenient to operate.
- LOW POWER CONSUMPTION-- The power consumption of the product is very low, saving money. The theoretical power consumption is less than or equal to 2 W.
- HIGH STABILITY-- The kilowatt hour meter has stable performance and can run steadily in the range of -20-55 degrees Celsius.
- APPLICATION-- 3 phase electric power meter can be widely used in power metering systems, standard: IEC62052-11 IEC62053-21.
Understand PUE figures in context
| Figure | Source and date | How to use it |
|---|---|---|
| 1.6 average PUE | U.S. Department of Energy/Federal Energy Management Program (DOE/FEMP), Best Practices Guide, 2024 | Guide-era context for an average data center, not a universal current target. |
| Below 1.1 | DOE/FEMP Best Practices Guide, 2024 | Described as achieved by some recent super-efficient centers; not a guarantee for every climate or AI load. |
| 1.55 annual average for large data centers in 2022 | Uptime Institute 2022 Global Data Center Survey, cited by DOE/FEMP in 2024 | A dated large-facility survey result with its own population and methodology. |
| 2.0 average; approaching 1.0 for highly efficient facilities | DOE/FEMP Cooling Water Efficiency Opportunities for Federal Data Centers, January 9, 2019 | An older page using a different guide context; do not merge it with the 2024 figures as if they were one benchmark. |
The theoretical lower limit of PUE is 1.0 because facility energy cannot be less than the IT energy counted in the denominator. Real sites vary with climate, redundancy, utilization, operating policy and boundary choices.
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 minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Define “tokens per watt” before measuring it
Watt is a rate (joules per second), while tokens are an amount of output. The ratio becomes meaningful only when its time basis is explicit. Measure energy over the run in joules or kilowatt-hours, and report throughput and power separately:
- Tokens per joule: defined output tokens divided by measured joules.
- Tokens per watt-hour: defined output tokens divided by measured watt-hours. Convert from joules using 1 Wh = 3,600 J.
- Throughput: output tokens per second.
- Average power: joules per second over the same measurement window.
Throughput divided by average power has units of tokens per joule. Calling that result “tokens per watt” without stating the time basis invites incorrect comparisons.
Fix the workload definition
Record all of the following in the benchmark description:
- Model name, exact version or checkpoint, quantization and context-window settings.
- Accelerator model and count, host CPU, memory, interconnect, firmware and driver versions.
- Inference software, runtime, compiler, kernel, precision, batching, concurrency, sequence length and generation limits.
- Prompt set, input-token accounting, output-token accounting and whether prompt processing is included.
- Warm-up policy, cache state, start and stop timestamps, and the number of repetitions.
- Quality or task-completion criterion, refusal handling, maximum latency and any output truncation rule.
- Power boundary: accelerator only, complete server, rack, IT system or whole facility. Identify whether cooling and power-conversion losses are inside it.
A higher token rate can result from larger batches, shorter answers, lower precision, a different model or a weaker quality threshold. It is not automatically a more efficient service.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Rank #4
- LCD Display Screen: The electric meter uses digital backlight LCD display which shows three-phase voltage, current, power, frequency and other electrical parameters display, active and reactive energy display.
- RS485 Communication: With a RS485 communication interface to host computer for data exchange, the digital power meter comply with the Modbus-RTU protocols, the terminals are marked A and B.
- Real Time Alarm: The three-phase power meter can do real-time alarm of loss of press, over-voltage, under-voltage, phase failure, loss of flow, over-current, current cutoff, overload and so on.
- Harmonic Measurement: Intelligent Power Meter can measure 2-31 harmonics of each phase voltage /current(graded harmonic content, total distortion rate, odd total distortion rate, even total distortion rate).
- Wide Applications: This meter is applicable to campus dormitories and large-scale of electrical energy metering, can also be used in enterprises and institutions for power management assessment.
Use a reproducible measurement window
- Load the exact model and software versions, then allow a documented warm-up period until temperatures, clocks and power stabilize.
- Run a fixed prompt and response set at the chosen concurrency or batch size. Keep scheduling, sampling and stopping parameters unchanged between comparisons.
- Start synchronized logs for token counts, completion status, accelerator utilization, clocks, temperatures, throttling flags and electrical energy.
- Stop after the predefined workload completes or the fixed interval ends. Record rejected, timed-out and failed requests rather than silently dropping them.
- Calculate total defined output tokens, total energy, average power and tokens per second. Report the quality or completion rate beside the ratio.
- Repeat under representative load and ambient conditions. Provide meter accuracy or uncertainty when available.
A concise result statement can follow this pattern in prose: for a named model and hardware/software configuration, a stated number of defined output tokens were generated during a stated interval using a stated amount of energy measured at a named boundary and meeting a named quality criterion. That sentence is the minimum needed for another team to understand what the number means.
Keep cooling and workload tests synchronized
Cooling is not merely a facility overhead. Lawrence Berkeley National Laboratory reports that thermal status can affect hardware performance, and that direct-to-chip liquid cooling can keep chips below thermal-throttling thresholds and change completion time. The same job can therefore produce different throughput under different cooling conditions.
For a cooling comparison, hold model, prompts, software, batch size and quality constant. At the same time, record rack-inlet temperature, accelerator temperature, clock frequency, utilization, throttling events, fan speed, cooling mode and outdoor conditions. Compare both facility energy and workload output. A design that lowers cooling electricity but causes throttling may deliver fewer completed tokens; a design that raises pump energy but prevents throttling may finish the workload sooner.
Repeat tests across low, medium and high utilization. Economizer hours depend on climate, set points and available outdoor conditions, so one cool-day run cannot represent a full year. Avoid overly low temperature set points and unnecessarily narrow humidity bands: DOE/FEMP operating guidance identifies both as sources of avoidable cooling demand. Hot-aisle/cold-aisle management should prevent mixing between cool intake air and hot exhaust.
Free tools Windows power users keep installed
One-click scans. No signup required.
Add water and carbon metrics when they matter
Energy-only reporting can conceal important trade-offs. ISO/IEC 30134-9:2022 covers Water Usage Effectiveness (WUE) during the use phase. DOE/FEMP presents a site-based form as annual site water use in liters divided by annual IT energy in kilowatt-hours. State whether water includes potable, reclaimed, blowdown or other sources and keep the time period aligned with the energy denominator.
Best Value
- 3-Phase Power Meter With Real-Time Data: see energy usage and solar generation vs. consumption with instant readings in kWh, Watts, Amps, Volts, Power Factor and cost ($).
- Connection Options- by WiFi (2.4 GHz) or wired Ethernet. NEED HELP? Reach out to Eyedro's Technical Support Team for assistance with the setup of your Eyedro device.
- High Resolution Watt Meter Data: once connected, Eyedro maintains full 24/7 usage history. Supporting fixed, tiered and time-of-use rate structures, MyEyedro features responsive real-time graphs, providing you with hourly, daily, weekly and monthly consumption and costs, as well as estimates of what your current electrical bill will cost.
- Multiple Voltage Options: 600V Delta, 480V/277V Wye, 208V/120V, 3-Phase High-Leg Delta and more.
- Directional Solar & Net Metering Ready: differentiates power consumed vs. generated, perfect for solar, grid-tie and net metering scenarios. To break out grid consumption from solar generation and use the MyEyedro Net Meter plugin 2 x Eyedro monitors are required.
ISO/IEC 30134-8:2022 covers Carbon Usage Effectiveness (CUE), operational carbon emissions intensity during the use phase. Disclose the emissions factors, market or location-based method and treatment of on-site generation. A liquid-cooling change can lower energy while increasing water use, or reduce water while shifting electricity demand; PUE, CER, WUE and CUE should remain distinct.
Instrumentation and data architecture
Electrical measurements
Place revenue-grade or appropriately accurate meters at the boundary named in the report. For three-phase plant equipment, use a three-phase power meter or analyzer matched to the site’s voltage, current transformers, phase configuration and safety requirements. Capture real power, reactive power, voltage, current, power factor and cumulative energy where available. Use the same integration interval for every meter and align clocks before analysis.
Thermal and environmental measurements
Use calibrated temperature sensors at supply and return points, flow meters suited to the coolant, and rack-inlet sensors at representative heights and aisles. Log humidity, outdoor dry-bulb and wet-bulb conditions, valve positions, pump speed, fan speed, chiller lift and cooling mode. Document sensor placement; a single outlet sensor cannot represent a high-density AI row with uneven airflow.
Software and telemetry
Collect accelerator power and temperature from the vendor telemetry interface, but do not assume it equals wall energy. Host CPUs, memory, networking, storage, voltage-regulator losses and server fans may be outside an accelerator-only reading. Reconcile telemetry with an external meter at the chosen server or IT boundary.
Compare systems on five axes
| Axis | What to hold or disclose | Why it changes the conclusion |
|---|---|---|
| Cooling energy and thermal state | CER boundary, pump/fan/chiller loads, temperatures, throttling | Thermal limits can change completion time and output. |
| IT energy | Accelerator, server, rack or IT boundary; meter accuracy | A narrower boundary generally reports a lower denominator than whole-IT energy. |
| Workload output | Model, prompts, input/output tokens, batch, concurrency and quality | Token rate is workload-dependent and can be raised by changing task difficulty. |
| Resource impacts | WUE and CUE methods, factors and periods | Energy, water and carbon can move in different directions. |
| Time and conditions | Run duration, annual versus interim label, ambient conditions | Short tests and favorable weather are not annual performance. |
Common measurement mistakes and fixes
- Calling PUE a cooling metric: publish CER for cooling and PUE for total facility overhead.
- Reporting an instantaneous watt reading: integrate energy over the workload window and give average power and throughput separately.
- Omitting prompt processing: state whether input tokens and prefill energy are included; use the same rule in every comparison.
- Changing quality to improve the ratio: keep a fixed task-completion or quality threshold and disclose refusals, truncation and failures.
- Ignoring embedded fans: identify whether server fan modules are counted as IT or cooling energy. LBNL notes that total PUE can differ from measured PUE because of such embedded equipment.
- Comparing unlike periods: do not compare a one-hour tokens-per-energy run with an annual PUE as though they were equivalent indicators.
- Using a boundary that moves between cases: draw the boundary once and apply it identically to every design.
- Treating dated averages as targets: attach the source, year and population to every benchmark figure.
- Assuming CER covers non-electric cooling: disclose heat-driven absorption chillers separately because the ISO/IEC 30134-7:2023 scope excludes them.
A reporting checklist
Before publishing or approving a result, verify that the record contains:
- Metric name and standard edition, if one applies.
- Facility, IT, cooling and workload boundaries shown in words or a diagram.
- Meter models or accuracy class, calibration status and sampling interval.
- Measurement dates, duration, annual/interim label and ambient conditions.
- Cooling equipment included, excluded or assigned to another boundary.
- Model version, hardware, software, prompt set, token-counting rule and quality criterion.
- Batch size, concurrency, warm-up, repetitions, failures and stopping conditions.
- Total energy in joules or kilowatt-hours, output tokens, throughput and average power.
- Temperature, utilization, clocks and throttling evidence for AI runs.
- WUE and CUE methods when water or carbon is decision-relevant.
- Uncertainty, missing data, maintenance periods and any deviations from the test plan.
How to interpret the final numbers
Use CER to discuss cooling-system energy performance, PUE to discuss facility overhead, and tokens per joule or tokens per watt-hour to discuss a defined AI service. Never rank facilities from one metric alone. A site can have a favorable PUE but poor model efficiency, or excellent token productivity at the cost of high water use. The defensible conclusion is the one that keeps boundaries, workloads, quality, thermal state and time period visible beside every number.
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.
