I doubt it. I think the big take away lesson from the language specific machines was that it's not worth it. LOTS of money is spent on the R&D that goes into making x86, ARM, etc. faster and better each iteration, and without the economies of scale that come with a huge install base, it's basically just a money pit. Even if a dedicated Mathematica CPU was faster now, the x86 version will be faster in a few years anyway.

It might make sense one day when when processor improvements level off even more than they already have.

On the other hand, it might be a neat idea to OEM a high end x86 system and sell it as a purpose built "mathematica machine," with Mathematica pre-installed, the best supported GPU, optimal RAM and CPU, etc.

A lot of Mathematica is already written in Mathematica. When Apple switched Macs to Intel CPUs, they included Mathematica as a flagship program, and flew their engineers out to do a port in secret. One engineer said the hardest part of the port was deciding what to do with the rest of the weekend.

How much was because a lot of it was written in mathematica vs. it already being cross platform an running on x86 on Windows/Linux?

source?

WWDC 2005: https://www.youtube.com/watch?v=kcfGsOKXO5M

You would not really need a specialized CPU anymore for a Lisp Machine. Those CPUs were invented to keep instructions small and efficient, provide large address spaces and enable simple compilers. In the last part of the 80s the trend to RISC CPUs was clear and several RISC CPUs were developed for Lisp, but they did not reach the market anymore. The SPARC cpu was mildly inspired by Lisp / Smalltalk. Since new CPUs were not happening for Lisp (there was no money for it), two of the existing CPUs were ported as VMs to other processors (Symbolics Ivory to ALPHA because of its 64bit support) and the Xerox Lisp system was ported to x86 and SPARC.

Besides the CPU there is another, even more important, difference. Lisp systems were all the way down to the metal developed in Lisp. Mathematica is largely written in C++ (runtime, environment, ...) and runs on top of a conventional OS (Windows, OSX, Linux, ...).

> I think the big take away lesson from the language specific machines was that it's not worth it.

I think the take away was learned when every generation of the x86 would be a significant leap in performance. As you say in paragraph 3, things might change with the leveling off of improvements.

> On the other hand, it might be a neat idea to OEM a high end x86 system and sell it as a purpose built "mathematica machine," with Mathematica pre-installed, the best supported GPU, optimal RAM and CPU, etc.

I've been pitched by salesmen that their machine is a killer machine for Mathematica. We don't use it here, so I was not really into it as much as buying the sealed medical computers for the carpentry people.

And what good would that be? Besides being single purpose and not running tons of other software?

It just seems like they are including everything into their product that I could actually see them selling a dedicated device. I could almost see them building an actual notebook (well, tablet) that runs only Mathematica.

Lisp Machines had already something like Mathematica.

It's called Macsyma and existed long before Mathematica. Macsyma was ported to Lisp Machines, in fact Lisp Machines exist partly because of Macsyma. It was one of the early Lisp applications which needed better and dedicated (not timeshared) hardware.

http://lispm.de/macsyma/macsyma.html