Corrigendum for February 2024 Research Article

A detailed explanation from the authors on the errors they caught and corrected in their February 2024 research article.

'Corrigendum' text on image of overcrowded adapters in power outlet, illustration
In the February 2024 Communications article titled, “Energy and Emissions of Machine Learning on Smartphones vs. the Cloud,” the authors found and corrected two arithmetic errors after it was printed. The corrections affected numerous derived values in the tables and the discussion, but the magnitude of the changes was not sufficient to affect the paper’s conclusions. Therefore, ACM published a Corrected Version of Record (CVoR) on January 25, 2024. For reference purposes, the VoR may still be accessed via the Supplemental Material section in the ACM Digital Library. The following is a detailed explanation from the authors on how the errors occurred and the ramifications on their findings.

Our paper, “Energy and Emissions of Machine Learning on Smartphones vs. the Cloud”on page 86 [in the February 2024 issue of Communications,] had two arithmetic errors that we caught shortly before the online publication but too late for the paper magazine. They affected some numbers, but did not change the qualitative conclusions: training on smartphone can be two orders of magnitude of the CO2e of training in the Cloud; the vast majority of smartphone energy use and CO2e comes from the chargers rather than the phones themselves; and the embodied CO2e from manufacturing computers in 2021 was about two orders of magnitude larger than the operational CO2e from ML in smartphones and Google data centers. Table 1 overestimated the Charger PUE by a third due to an error in our logic. The reason is that we multiplied the average number of chargers per phone (2.7) times average energy use per charger, when the formula should have been the average energy use of 1 charger plus 1.7 times the average vampire energy use of chargers. This reduction of Charger PUE also percolated through the CPUE sensitivity numbers in Table 2. This second error is far less subtle, miscalculating CO2e when simply converting from watt-hours of energy into CO2e. Figure 3 shows embodied CO2e from manufacturing computers in 2021 was 160 megatons versus originally 1.2 megatons for ML in smartphones and Google data centers, but the correct number for the latter should have been 2.0 megatons. As far as we can tell, the mistake occurred when an early draft increased the energy use of ML on smartphones in one paragraph but the corresponding CO2e was left unchanged. Despite seven co-authors, many rounds of internal reviews, and two rounds of CACM reviews of three versions of the paper in 11 months, this relatively obvious blunder went unnoticed until recently. Besides changes to these two tables and one figure, we had to correct some of the summary statements in the introduction and the conclusion and the CO2e numbers for smartphones in the bottom part of the last column of page 94. We regret and apologize for our mistakes and thank the CACM staff and editors for letting us fix them.

Join the Discussion (0)

Become a Member or Sign In to Post a Comment

The Latest from CACM

Shape the Future of Computing

ACM encourages its members to take a direct hand in shaping the future of the association. There are more ways than ever to get involved.

Get Involved

Communications of the ACM (CACM) is now a fully Open Access publication.

By opening CACM to the world, we hope to increase engagement among the broader computer science community and encourage non-members to discover the rich resources ACM has to offer.

Learn More