There's an interesting design point in high wattage processors. If your bytecode interpreter fits comfortably in the L1 cache, and the bytecode is denser than machine code, and the program doesn't fit in the caches, then there's potential for interpreted bytecode to outperform native code. Quirk of really fast compute paired with moderately fast memory. I can see that working particularly well if one interpreter instance executes code for multiple processes.
