Interesting approach for extracting ILP -- though I guess it's been tried before (first commenter to name Transmeta wins a...a rubber duck!) and it doesn't seem to have caught mainstream attention. The forecasted end of Moore's Law may change things there, though.
What immediately came to mind was the optimization overheads. Although the optimization cache (the 128MB chunk of main memory) should ameliorate the overhead for frequently used applications, infrequent applications or those that show very input-dependent behavior may not get as large benefits.
Also, I expect benchmarking this and comparing against the competition will be non-trivial.