💻 computer science

Power Consumption Patterns Using Telemetry Data

This paper analyzes Intel telemetry data to demonstrate that user behavior, rather than just hardware choice, significantly influences power consumption patterns, offering insights for stakeholders to reduce environmental impact through targeted improvements.

Original authors: Harry Cheon, Yuyang Pang, Zhiting Hu, Benjamin Smarr, Julien Sebot, Bijan Arbab, Ahmed Shams

Published 2026-02-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Harry Cheon, Yuyang Pang, Zhiting Hu, Benjamin Smarr, Julien Sebot, Bijan Arbab, Ahmed Shams

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

🌍 The Big Idea: It's Not Just the Car, It's the Driver

Imagine you buy a new car. You might think, "This car is a gas-guzzler because it has a huge V8 engine." That's the hardware. But what if your neighbor drives the exact same car, yet they use half as much gas? Why? Because you drive aggressively, speed up constantly, and leave the engine running while you grab coffee. Your neighbor drives gently and turns the engine off when stopped. That's the user behavior.

This paper is about a team of researchers (from UC San Diego and Intel) who looked at millions of computers to figure out: What actually makes a computer use more electricity?

For a long time, Intel and others thought, "It's all about the hardware. If we make better chips, we save power." This paper says, "Hold on. The way people use their computers matters way more than the computer itself."

🔍 The Investigation: The "Telemetry" Detective Work

Intel has a website where millions of people go to update their drivers (the software that helps hardware talk to the computer). When they do this, they can choose to share "telemetry" data. Think of this as a black box on an airplane. It doesn't record your name or your home address (privacy is key!), but it does record:

How hot the CPU gets.
How much power it's using.
What apps you are running.
How long you keep the screen on.

The researchers looked at a sample of 1 million computers from around the world. They wanted to see if the "driver" (the user) mattered more than the "car" (the hardware).

🇨🇳 vs. 🇺🇸: The Great Power Divide

The most interesting part of the study was comparing China and the United States. These two countries are the biggest players in the world for both computer usage and carbon emissions.

The Finding:
The average computer in China uses almost twice as much electricity as the average computer in the US.

The "Why" Mystery:
The researchers asked: Is it because Chinese people have bigger, more powerful computers?

Answer: No. Even when they compared computers with the exact same specs (like a 15-watt laptop vs. a 15-watt laptop), Chinese users were still using way more power.

The "Driver" Habits:
So, what are the Chinese users doing differently?

The Night Owl Effect: In the US, people tend to use their computers heavily in the morning and then wind down. In China, power usage actually rises as the day goes on, peaking late at night.
The Weekend Dip: Americans take a break on weekends; their computer power usage drops. Chinese users keep the power flowing steadily, 7 days a week.
The "Turbo" Trap: Computers have a "Turbo" mode (like a sports mode in a car) that uses extra power for speed. The researchers thought, "Maybe Chinese users just have more powerful computers." But even when the computers were in Turbo mode, Chinese users were still using more power. It turns out, they are running more intense tasks for longer periods.

🎮 The Surprising Culprits

The researchers built a "math model" (a linear model) to predict power usage. They expected hardware specs to be the biggest factor. Instead, they found that user behavior was the star player.

Here are the top things that make a computer hungry for electricity:

Gaming: (Obviously, this is a power hog).
Video Editing & Development: Coding and editing videos burn energy.
Web Browsing: Surprisingly, even just browsing the web matters.

The "Bad App" Discovery:
The researchers found a specific reason why Chinese users were using so much extra power. They discovered that popular Chinese video streaming sites (like DouYu and Bilibili) were not using a special "energy-saving" feature built into Intel chips.

The Analogy: Imagine you are watching a movie. The "energy-saving feature" is like a translator who speaks the movie's language perfectly. If the website doesn't use this translator, the computer has to work 10x harder to understand the video, burning extra fuel.
Because these popular sites weren't using the efficient tech, millions of computers were burning extra energy unnecessarily.

💡 The Takeaway: What Can We Do?

The paper concludes that we can't just wait for better hardware to save the planet. We have to change how we use our devices.

For Companies (like Intel): They need to talk to software developers. "Hey, if you use this specific energy-saving feature, your users will save a ton of power."
For Users: It's not just about buying a "green" laptop. It's about what you do on it. If you are running heavy software inefficiently, you are the one creating the carbon footprint.
The Future: The researchers hope to create a "Green Mode" for computers. Imagine a setting that automatically optimizes your apps to use less power, kind of like how a hybrid car switches between gas and electric to save fuel.

🏁 Summary in One Sentence

You can't just blame the car for high gas prices; sometimes, the driver's habits are the real reason the tank is empty, and fixing the driver's behavior saves more fuel than building a new engine.

1. Problem Statement

The paper addresses the challenge of reducing the global carbon footprint associated with computing devices. While previous discourse often attributes high power consumption primarily to hardware choices (e.g., processor type, device class), this study challenges that prevailing sentiment. The authors posit that user behavior is a more significant determinant of power consumption than hardware selection.

The specific problem is to analyze Intel's telemetry data to:

Understand global power consumption patterns.
Identify the drivers of significant disparities in power usage between major markets (specifically the US and China).
Determine if optimizing user workflows and software usage can yield greater energy savings than hardware optimization alone.

2. Methodology

Data Source and Collection

Dataset: The study utilizes a 1-million GUID (Global Unique Identifier) sample from Intel's Intel Driver & Support Assistant (IDSA) telemetry data.
Privacy: Data is anonymized; no PII (Personally Identifiable Information) such as IP addresses, MAC addresses, or serial numbers is collected. A GUID stitches daily data points to a specific device over time without revealing user identity.
Scope: Approximately 40% of IDSA users opt-in to telemetry. The dataset includes over 1,400 parameters per device, including CPU power, temperature, network traffic, application usage duration, and frame rates.
Granularity: Data is primarily aggregated daily (24-hour cycles) due to consistency issues with finer time intervals.

Data Processing and Metrics

Normalization:
- Weighted Mean: Used to aggregate multiple entries for the same day, weighted by sample count.
- Normalized CPU Usage ( $\eta$ ): Calculated as $\eta = c \times \omega_u$ , where $c$ is the number of cores and $\omega_u$ is the weighted mean CPU utilization. This normalizes workload across different core counts (e.g., equating 25% usage on a 4-core to a single-core equivalent).
Key Metrics:
- MUAP (Median User Average Power): The median of the arithmetic mean daily power consumption for users within a country. This represents the "typical" user's power usage.
- Turbo Status: A binary classification where a machine is considered "turbo-ed" if it spent >10% of its active time in Turbo Boost mode.
- Work Intensity: Analyzed by binning normalized CPU usage and comparing power distributions across bins.

Analytical Approach

Exploratory Data Analysis (EDA): Comparative analysis of the US and China across multiple dimensions: time of day, day of week, OEM (Original Equipment Manufacturer), turbo status, and work intensity.
Linear Modeling:
- Model Type: LASSO (Least Absolute Shrinkage and Selection Operator) and Linear Regression.
- Features: 225 parameters spanning 5 tables (Software Usage, Website Usage, C-State, AC/DC Info, System Info, CPU Package Power).
- Validation: 80/20 train/test split. Performance measured via Mean Squared Error (MSE) and $R^2$ .
- Goal: To identify which features (behavioral vs. hardware) have the highest coefficients in predicting CPU package power.

3. Key Contributions

Shift in Paradigm: The paper provides empirical evidence that user behavior is a more critical factor in power consumption than hardware choice, challenging the industry assumption that hardware selection is the primary determinant.
Cross-Cultural Behavioral Analysis: It reveals distinct power consumption patterns between the US and China that persist even after controlling for hardware specifications (TDP and CPU type).
Software-Specific Insights: The study identifies specific software categories and platforms (e.g., streaming sites like DouYu and Bilibili) that drive excessive power consumption due to inefficient codec usage or lack of hardware acceleration.
Actionable Framework: It proposes a methodology for stakeholders (like Intel) to target software optimization and user education to reduce carbon footprints effectively.

4. Key Results

Global and Regional Disparities

MUAP Gap: China's Median User Average Power (11.74W) is nearly double that of the US (6.4W).
Persistence: This gap remains significant even when controlling for CPU Thermal Design Power (TDP) and device type (laptop vs. desktop).
OEM Consistency: The higher power consumption in China is observed across almost all top OEMs (e.g., HP, Dell, Lenovo), with HP showing the largest disparity (6.47W difference).

Temporal and Behavioral Patterns

Time of Day:
- US: Power peaks in the morning (~8 AM) and declines throughout the day.
- China: Power consumption increases as the day progresses, peaking at night.
Day of Week:
- US: Clear "weekend dip" in energy consumption (Saturday/Sunday).
- China: Consistent energy consumption throughout the week, with no significant weekend reduction.
Turbo Usage: Contrary to expectations, the power distribution gap between the US and China is larger when machines are in Turbo mode. This suggests that Chinese users not only use Turbo more frequently but also sustain high-power states longer.

Linear Model Findings

Model Performance: The LASSO model achieved an $R^2$ of 0.33, which is significant given the complexity of the system and unmeasured variables.
Feature Importance: The top 5 features with the highest absolute coefficients were:
1. CPU Normalized Usage (C-state)
2. Number of Cores
3. Persona: Gamer (User Behavior)
4. CPU Family: Core i7
5. Persona: Web User (User Behavior)
- Insight: Three of the top five predictors are related to user behavior (Gaming, Web usage) rather than raw hardware specs.
Software Impact:
- Gaming is the highest energy consumer.
- Web-based applications generally show lower energy coefficients, suggesting cloud computing is more efficient per operation.
- Inefficiency Discovery: Internal Intel investigations confirmed that popular Chinese streaming platforms (DouYu, Bilibili, IQIYI) do not utilize Intel's Si Hardware-based DRA (Dynamic Rate Adaptation) or efficient codecs, leading to significantly higher power draw compared to Western counterparts.

5. Significance and Implications

Environmental Impact: Since the US and China are the top two CO2 emitters, understanding and mitigating the behavioral drivers of power consumption in these regions offers the highest potential for global carbon reduction.
Strategic Recommendations for Intel:
- Software Optimization: Collaborate with software vendors (especially streaming and video platforms) to implement hardware-accelerated codecs and DRA.
- User Education: Promote web-based applications over local heavy processing where feasible, leveraging the efficiency of cloud data centers.
- "Green Mode": The authors propose developing real-time power optimization modes based on user behavior patterns.
Future Work: The study suggests expanding the analysis to more regions, integrating regional carbon intensity data (emissions per kWh), and building more complex predictive models to refine the "Green Mode" concept.

In conclusion, the paper demonstrates that while hardware matters, how users interact with their devices is the dominant variable in power consumption. Addressing software inefficiencies and user habits offers a pragmatic and immediate path to reducing the computing industry's carbon footprint.