Retraction: our benchmark laptop was running at about one fifth speed

01Why it matters

A speed number is meaningless if the machine was secretly throttled.

02Method, step by step

1. Print the current core ratio against the requested ratio at boot.
2. Apply a fix, clearing bd-PROCHOT and requesting the maximum P-state.
3. Generate the same 214 tokens with two same-day builds.

03Results, exact and sourced

1. R1The CPU was clamped to ratio 5 of a requested 27, with PROCHOT asserted while the thermal sensor read cool.

   was clamped to core ratio 5 out of a requested 27.

2. R2Every earlier number from that box (before 2026-07-29) was taken at roughly a fifth of nominal clock, and was retracted.

   Every performance number we took on that box before 2026-07-29 was measured on a CPU

3. R3After the fix, cur_ratio went from 5 to 25 on 3 of 3 cold boots (one machine).

   Reproduced on 3/3 cold boots.

4. R4Wall-time ratio across the two same-day builds was 3.861x, below the 5.000x the ratio change implies; output was byte-identical. Not reported as a speedup.

   We are deliberately not reporting this as a speedup.

5. R5Published as a retraction.

   this post is a retraction, not a speedup

6. R6Limitation of the before/after: no clear-only arm.

   No clear-only arm.

05Scope and caveats

• One machine (n=1 laptop, 3 cold boots); run by us; no unaffiliated reproduction.
• The fix (MSR 0x1FC bit 0) is about fifteen years old; the post cites prior art and does not claim it as new. The measurement error is the point.
• The fix stage changed two things at once (cleared bd-PROCHOT and requested the maximum P-state); no clear-only arm was run, so the ordering hypothesis is untested.
• Durability is untested: unknown whether the clear survives S3 suspend or an EC re-assert after a firmware event.
• The 3.861x vs 5.000x gap is described as a crude indicator of memory-bound work, not a measured decomposition.
• The post's tok/s column is too coarse to quote; tick counts are the instrument.

06Outcome

07What we learned

Measure the machine before measuring the code. We retracted every earlier number from that box instead of reporting a speed-up.

08Check it yourself

• alice-aegis / 2026-08-05_bd-prochot-benchmark-lie.mdhttps://github.com/Aefinity-AI/alice-aegis/blob/main/docs/posts/2026-08-05_bd-prochot-benchmark-lie.md

Read the post (link above).