This is why Chuck Moore designed his own chips like GreenArrays I believe. There is no impedance mismatch between software and hardware and the entire thing is optimized for the problem domain and massively parallel. I'm sure you can get more speed out of C, but it no other language (even Lisp or Smalltalk) can the user understand the entirety of the system. Lisp machines and Smalltalk machines probably are about as close as possible, but I have zero experience with them personally. Perhaps the project to put PicoLisp on bare-metal qualifies if one uses a simple enough microcontroller as the language itself is about as simple as you can get a practical lisp.

Over the last twenty years or so, I have written serious Forth code generators for six CPU architectures. There's only one code generation algorithm that's not regularly present in most Pasgol family compilers. The results are the VFX Forth systems at http://www.mpeforth.com.

I've looked over your products over the years and been super impressed by your results. Thanks for all your hard work!

Not really, modern Lisps have done a fine job optimizing performance. FORTH suffers from treating the hardware too granularly, and thus not being able to take advantage of any of the advances in hardware in the last 2 decades.

There is no a priori reason FORTH can't take advantage of modern hardware, since in the extreme case the compiler can be completely replaced by redefining the meaning of ':' and ';'.