It's lock-free because it uses ordered loads and stores, which is also how you implement locks. I find the semantic distinction unconvincing. The post is really about how slow the default STL mutex implementation is.