Just like how software transactional (STM) memory will be for morons once the average desktop computer gets around 16 cores.
Real programmers manage their own memory and are aware of the memory footprints of their data structures as well as hardware cache coherency and memory hazards. The build their own memory allocators to suite different requirements.
Just like how real programmers today use fine-grained lock-free and wait-free synchronization primitives such as multi-producer/single-consumer bounded queues and RCU (read-copy-update) and build their work/task schedulers to be NUMA aware. You loose all of that fine grained control with one-size fits all STM.
This is why C/C++ will continue to be the language of choice for high-performance computing even in the age of massively parallel computation. C1x/C++0x have well defined memory models around which atomic operations can be implemented.
Enjoy your garbage collected and sequentially consistent programming environments, morons.