Longpoll is currently not implemented in the bfgminer getwork proxy, so that will result in low efficiency / lots of stales.  Using slush's proxy, cubes do appear as separate workers to the pool when running a single instance of slush's proxy, however, I can't recommend that either.  While playing around with slush's proxy and the stratum-mining pool software, I discovered that if you have multiple cubes connected to one instance of slush's proxy, then the proxy will get the share difficulties confused and only submit shares that meet the highest difficulty requirements of all workers.
Based on the above findings, currently the best setup is ONE INSTANCE OF SLUSH'S PROXY PER CUBE, however much cpu time that ends up using.  I really wish longpoll would be implemented in the bfgminer getwork proxy soon...  Either that, or the cube firmware should be updated to support stratum natively.