A first-person correction to a common story about strong and weak CPU memory models: modern cores usually speculate and detect ordering violations later, so the practical cost is concentrated around contention rather than every access.
It replaces a slogan with a more useful performance model: reduce shared contention first, and treat memory-model strength as one implementation trade-off rather than a scalability verdict.
- computer-architecture
- concurrency
- performance
Why was this selected?
- Heading: Does the linked post use this exact title?
- Description: Does the article explain speculation, contention, and recovery instead of treating strong ordering as literal serialization?
- Why Read: Does the author end with a practical lesson about reducing contention?
Reviewer: “A rare, expert correction of a widely repeated claim about memory models: CPUs speculate and roll back on ordering violations, so the real cost concentrates on contention. Ends in the practical 'contend less' mantra rather than an architecture verdict.” Caveat: Verified against full entry text; the author's empirical claims about scalability differences are his own experience, not measured data.