Power Consumption by Country

REPO: https://github.com/harrycheon/intel-cf-public/tree/main

Notes: By Trey

Goal: reduce carbon emissions by reducing energy consumption by identifying where intel should improve their products. Intel’s current belief: hardware determines energy consumption New Findings: user behavior is more important than hardware (can Intel adapt for that?)

Why split by country?

GUID is an ID for a device(or “machine”) constructed without recording IP / MAC addresses / emails etc. Is GUID safe… why do we need DP? -reconstruction and linkage using device properties can expose identity of GUID.

Database Tables used Daily aggregates from each device • hw_pack_run_avg_power: Average packagepower consumption for collection period • os_c_state: CPU Consumption • sampler_data: CPU turbo data • sysinfo: System meta information such as country, type of machine \ (laptop or desktop), type of CPU, Original Equipment \ Manufacturer(OEM) and more • web_cat_usage_v2: Website usage information, including browser types, content \ categories, and duration • frgnd_backgrnd_apps_v4_hist: Software usage information, contains software names, \ duration, AC/DCstatus, display ON/OFF, and more

Random Sample of 10,000 GUID’s and over time 1 million rows total (avg 100 days of data per device)

MSE 80:20 train test Results: The LASSO model yields an RMSE* of 5.86 (watts) on the test set. The R2 for the linear regression model is 0.33. (Interpretation: Given that many factors potentially influencing the CPU package power aren’t captured in the telemetry data or weren’t included in our dataset, an R2 value of 0.33 is notably significant.) *Author said it was rmse

”the results suggest that the majority of the difference in distribution between the US and China is driven by users that turbo more often.”

Why would intel push for web based apps? In general lowering usage patterns would mean using their product less…

”total duration spent on each category. Identifying the software categories that are the most energy-intensive enables Intel to prioritize the optimization of applications within these categories and design CPUs with enhanced energy efficiency specifically tailored for these applications.”

Lasso predicts power_mean: SUM(pw.nrs * pw.mean) / SUM(pw.nrs)

nrs: Number of samples (sampling frequency most commonly 5 seconds)

mean: the mean of the metric

pw from hw_pack_run_avg_pwer from HWIL (hardware input library)

50 Watts is normal for U-series CPU (GPU’s ~150-1500)

watt = Joule / sec

“Simply, energy is capacity to do work; power is the rate at which work can be done. Power (watts) is applied over a period of time (hours), giving rise to a figure in watt-hours that shows how much energy has been consumed.” Wh = Watts * hours

power data from 01/2019-09/2022

privately calculating MSE and R^2 protected by post processing on y_pred

LASSO → l1 ball constraint → 2p sides of the constraint set = 2p vertices which is polynomial in p which makes it a polytope. → DP bounds

Implementation

capstone-env-1