I was under the impression that STM, being an opportunistic (success biased) concurrency primitive, was faster than explicit locking. Isn't it more performant?

I was under the impression that STM, being an opportunistic (success biased) concurrency primitive, was faster than explicit locking. Isn't it more performant? I remember I saw in a talk given where in that benchmark STM was as fast as no locking at all !