Software transactional memory (STM) is a promising technique for writing concurrent programs. So far, most STM approaches have been experimentally evaluated with small-scale microbenchmarks. In this paper, we present several surprising results from implementing and experimenting with STMBench7 – a large scale benchmark for STMs. First, all STMs we used crashed, at some point or another, when running STMBench7. This was mainly due to memory management limitations. This means that, in practice, none of the tested STMs was truly unbounded and dynamic, which are the actual motivations for moving away from hardware transactional memories (HTM). Performance results we gathered also differ from previously published results. We found, for instance, that conflict detection and contention management have the biggest performance impact, way more than other aspects, like the choice of lock-based or obstruction-free implementation, as typically highlighted. Implementation of STMBench7 with various STMs also revealed several programming related issues such as the lack of support for external libraries and only partial support for object oriented features. These issues prove to be a major limitation when adapting STMs for production use. Our work is by no means a bashing of prior work on STMs. All STMs we considered are very well designed and implemented. What we highlight here is that providing genuinely unbounded transactions is a hard and complicated task, but full of interesting technical and research problems. Solutions to these problems should be evaluated against large scale benchmarks, like STMBench7.