There are many engineering obstacles when scaling to advanced processor nodes that must be surmounted, and every one counts. This one had had many designers worrying direct-write shared memroy caches would have to be replaced with scratchpad or message passing schemes. Luckily, SRC has shed some light on this issue, hopefully keeping designers from fixing a architectureal feature that is not going to break all the way out to 512 cores per processor chip.
Personal option: the issues of having cache dose not only result in storage overhead, but also introduce unpredictability e.g. in terms of performance. This is particularly true for the systems with real-time requirements. As programmers do not have to take care of data movement, maybe, another option to make progress on compiler development to manage data movement on scratchpad memory.