Another round of speculative-execution vulnerabilities

I think efficiency depends heavily on how much performance you want. See the image halfway down https://community.arm.com/arm-community-blogs/b/announcem... , which shows power vs performance curves for Cortex-A510 (in-order; speculative instruction fetch but not really speculative execution), Cortex-A715 (OoO; speculative execution; "balance of performance and efficiency"), and Cortex-X3 (OoO; highest performance).

They're curves because power is a non-linear function of frequency. There's some overlap where a lower-end CPU near its max frequency has worse power than a higher-end CPU at equal performance.

Draw a straight line through the origin and tangential to the purple (middle) curve. That should represent the optimal power/performance ratio for the Cortex-A715. By my rough measurements on this questionably-precise graph, the Cortex-A510 curve shows better power/performance than that when it's about 20%-60% of its max performance.

So if you're trying to optimise power/performance, and you're happy with <15% of the Cortex-A715's max performance - maybe your task doesn't need to complete quickly, or maybe you've got an embarrassingly parallel problem and can spread it over 6x as many cores with no extra overhead - then the Cortex-A510 seems worthwhile. But if you need even slightly more than that, and would have to drive the Cortex-A510 at a higher frequency, you'll get better efficiency *and* 3x better performance by switching to the Cortex-A715 at half its max frequency.