mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-05-06 01:28:35 -05:00
[GH-ISSUE #944] Mlsysbook_chapter12_8_1_PowerMeasurement #1640
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: github-starred/cs249r_book#1640
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @formlsysbookissue on GitHub (Aug 21, 2025).
Original GitHub issue: https://github.com/harvard-edge/cs249r_book/issues/944
Originally assigned to: @profvjreddi on GitHub.
In Section 12.8.1, there are two issues affecting the clarity and credibility of the discussion on power consumption:
Unreferenced claim in paragraph 3:
The paragraph states that memory-bounded inference tasks can account for up to 60% of power consumption. However, no citation is provided to support this figure (not a negligible rate). Given its specificity, a reference would help validate the claim.
Abrupt DVFS mention in paragraph 5:
Paragraph 5 introduces DVFS (Dynamic Voltage and Frequency Scaling) with no prior explanation or transition. The surrounding paragraphs (4 and 6) remain focused on shared infrastructure and cooling in the context of power consumption, making the DVFS reference feel disconnected from the rest of the discussion (if placed here, it has right to be discussed more).
This paragraph also includes another unreferenced claim — that DVFS can reduce power usage by up to 50%. Again, a citation would greatly improve the credibility of this figure.
12.8.2: Similar to 12.8.1, adding brief context and proper citations would help readers follow the content more easily and trust the data presented.
@profvjreddi commented on GitHub (Aug 22, 2025):
Thanks @formlsysbookissue for this awesome feedback. I will be sure to take a close look it in the coming few days and push an update!
@profvjreddi commented on GitHub (Aug 24, 2025):
Thanks for catching those issues. Here's what I fixed:
The 60% power claim: You're right, that needed a proper source. I found the actual Google study (Boroumand et al.) that shows it's actually 57.3% for TensorFlow Mobile data movement. Much better than just throwing around unsupported numbers.
The DVFS thing: Yeah, that was pretty jarring how it just appeared out of nowhere. I rewrote that section to flow better - now it connects DVFS to the infrastructure discussion instead of just dropping it in randomly.
The 50% DVFS claim: Another unsupported number, I agree! Added the Kim et al. paper that actually researches this stuff.
Section 12.8.2: Same issues there - added proper citations for the quantization claims and frequency scaling stuff. Also cleaned up the writing so it doesn't jump around as much.
Basically took all your feedback and made sure every claim has a real academic source backing it up. The sections read much better now and don't feel like they're just making stuff up.
Thanks for the great feedback - this kind of detailed review really helps make the book better! 🙏
@profvjreddi commented on GitHub (Aug 25, 2025):
@formlsysbookissue Thanks again for raising this! I’ve closed the issue since I’ve updated the content. You can review the latest merged version here: https://harvard-edge.github.io/cs249r_book_dev/
(note: this is the DEV version, not generally public).
If you’d like, I’d be happy to tag you on future edits to your suggestions before committing so that you can tell me if my fixes address your questions/concerns, but I didn’t want to assume on your time without checking first.
Feedback on whether this revision addresses your concern would be very welcome.