I understand that the plan is to have specific hardware to allow IoT devices to perform the PoW, but how will they know the state of the network in order to perform the MCMC? Wouldn't that require more resources?
Tip selection and PoW are completely different things, which do not necessarily have to happen at the same point.
Tip selection performs MCMC on the tangle to find two transactions (or rather their hashes) your new transaction (or transaction bundle) should approve. This step is not deliberately computationally more expensive than it has to be, so it can be run on full nodes (like it is today when you are using a light wallet, the tip selection runs on a full node).
PoW is some technique to deliberately spend some computation time on finding a partial hash collision, similar to what other cryptocurrencies use to secure the mining. This is deliberately computationally more expensive than it has to be (to make it easier, just reduce the MWM) and is there to make it harder for adversaries to introduce thousands (or tens of thousands) of transactions very quickly into the system.
There are rumors that the IOTA foundation is implementing ternary coprocessor (which, among other things, is expected to perform PoW computation more energy efficiently than current binary processors). This does not have any effect on tip selections, and this coprocessor can probably also used for other tasks that are more efficient in ternary. But keep in mind that this is speculation and there won't be any official statements about it from the IOTA foundation (as they claim that the technology they are working with is under NDA).