- The paper finds that user behavior and software explain CPU package power more strongly than hardware, with China's median user average power reaching 11.74 W versus 6.4 W in the United States.
- Using telemetry from roughly 1 million device-days, LASSO and linear regression identify normalized CPU usage, gaming, core count, and CPU family as leading predictors, achieving a test MSE of 5.86 W² and an R² of 0.33.
- The findings point to practical efficiency measures such as optimizing gaming workloads, adopting hardware codecs and display-refresh controls, and improving power-aware software for high-consumption applications.
Overview and motivation
This paper analyzes CPU package power consumption using large-scale telemetry collected through Intel's Driver and Support Assistance (IDSA) ecosystem, drawing on a 1 million GUID (device) sample from a global fleet of consumer machines. The authors' central claim is deliberately contrarian within Intel: contrary to the internal belief that hardware choice is the primary determinant of a device's power draw, they find that user behavior is the dominant factor. Their motivation is pragmatic carbon accounting — since the carbon intensity of a watt-hour is outside the control of individual users or a single company, the most tractable lever is reducing energy consumption itself, in line with the reduce–reuse–recycle hierarchy. The study adopts Wiedmann et al.'s definition of carbon footprint and focuses on package power as a measurable proxy.
Data and methods
The telemetry originates from devices that opt in (~40%) to sending usage data when updating drivers via IDSA. Over 1,400 device parameters are collected — package power, temperatures, network and disk traffic, application launch times and durations, C-state data, and turbo status — with no PII; a separately generated GUID stitches daily records to the same device without identity leakage. ESRV agents collect at regular intervals, aggregate every 24 hours, and upload when the machine is active. Four main tables support the exploratory analysis: hw_pack_run_avg_power, os_c_state, sampler_data, and sysinfo.
Several normalization choices matter for interpretation. Multiple same-day entries are combined via sample-weighted means, and CPU utilization is multiplied by core count so that utilization reflects single-core-equivalent workload rather than raw percentage. Cross-country comparisons use a Median User Average Power (MUAP) metric: per-user mean daily power, aggregated by country, then medianed. Turbo status is defined as being in turbo for more than 10% of usage time in a day. For hardware-controlled analyses, the authors restrict to Intel Core U-series processors (mostly 15 W TDP).
The modeling component uses a random sample of 10,000 GUIDs yielding ~1 million GUID-day records across 225 features spanning five tables (software usage, web usage, C-states, AC/DC status, system information). LASSO regression serves for feature selection, ordinary linear regression for interpretable coefficients, with an 80/20 train/test split and MSE on the test set.
Cross-country findings
The headline result is stark: China's MUAP is 11.74 W versus 6.4 W in the US — nearly double — and China leads all top-10 represented countries. The US and China are also the two most represented countries in the sample (13% and 7%), and the two largest actors in emissions, motivating the focus.
The authors then systematically test whether this gap is an artifact of hardware or user mix, and find that it is not:
| Analysis |
Finding |
| UAP distributions |
Gap persists after controlling for CPU TDP |
| Time of day |
China's consumption rises toward evening; US peaks ~8 am then declines |
| Day of week |
US shows a weekend dip; China is flat across the week |
| OEM comparison |
Positive China–US difference for nearly all top-10 OEMs; HP largest at +6.47 W |
| Turbo status |
US–China divergence is larger on turbo days, contrary to expectation |
| Work intensity |
At matched normalized CPU usage, Chinese users draw more power |
Several of these results contradict the authors' own priors. The gamer-persona hypothesis fails: gamer proportions are nearly identical (11.44% US vs. 10.54% China). The turbo result is the most counterintuitive, since turbo operation is hardware-limited; the authors concede this may be an artifact of the 10% daily threshold and daily aggregation, suspecting the residual non-turbo usage drives the difference. The OEM result is corroborated externally — HP had independently contacted Intel about overheating in Chinese machines — and the persistence of the gap for gaming OEMs (MSI, Razer, Alienware) further undercuts a demographic explanation.
Linear model results
The LASSO model achieves a test MSE of 5.86 W² and the linear regression an R² of 0.33. Given that the telemetry omits many power-relevant factors, the authors argue this R² is substantial — and, more importantly, the coefficient structure supports the behavioral hypothesis: three of the top five features by absolute coefficient are behavioral (normalized CPU usage/C-state, gamer persona, web-user persona), alongside core count and Core i7 CPU family.
| Feature |
Coefficient |
| CPU normalized usage (C state) |
1.24 |
| Number of cores |
0.58 |
| Persona: gamer |
0.42 |
| CPU family: Core i7 |
0.42 |
| Persona: web user |
−0.22 |
Ranking software-category coefficients yields an energy ranking in which gaming dominates both instantaneous power and total energy consumption, followed by development/programming and multimedia editing when weighted by usage duration. Notably, browser-based usage carries a negative association with power, which the authors connect to the superior per-operation efficiency of hyperscale data centers and their renewable-energy procurement; they suggest Intel could steer energy-conscious users toward web-based applications.
The findings triggered internal Intel investigations that produced a concrete mechanism: Chinese streaming platforms (DouYu, Bilibili, IQIYI) do not use Intel hardware-based display refresh adaptation or hardware codecs, consuming substantially more power than comparable sites on identical hardware — independent validation of the behavioral/software-driven account.
Limitations and open questions
The authors are explicit that most analyses rest on daily aggregates rather than fine-grained time series, a pragmatic constraint that limits causal resolution — most visibly in the turbo analysis, where the 10% daily threshold conflates intra-day usage regimes. The R² of 0.33 also implies that roughly two-thirds of power variance remains unexplained, and the MUAP comparison does not fully adjust for confounders such as desktop/laptop mix. The claimed behavioral effect rests on observational data; no intervention or controlled experiment is reported. Open questions the paper leaves include whether the US–China gap generalizes to other underrepresented regions, how region-specific carbon intensity per watt-hour would change the prioritization, and whether finer-grained (sub-daily) telemetry would resolve the turbo anomaly.
Conclusion
Using a million-device telemetry corpus, this paper shows that user behavior and software stack — not hardware selection — are the strongest levers on CPU package power, evidenced by a near-doubling of MUAP in China relative to the US that survives TDP, persona, and OEM controls, and by coefficient structure in an interpretable linear model. The work translated directly into internal Intel investigations identifying inefficient media platforms as a mechanism, and it frames concrete optimization targets (gaming workloads, codec adoption) for Intel's 2030 sustainability goals.