logoalt Hacker News

K0balt • today at 3:18 PM • 1 reply • view on HN

All things being equal, I could see this leading to 1-2 order of magnitude in compute efficiency over von Neumann / Harvard architectures, and may be ideally suited to in-memory compute, in particular.

If you can program it in rustish language with the intermediate representation, that makes everything portable to it without having to figure out the parallelism, potentially.

Very impressive and potentially revolutionary. Micron or Samsung needs to see this, they are having to reinvent compute anyway.


Replies

thesz • today at 3:57 PM

  > without having to figure out the parallelism...
Parallelism can be regular (matrix or tensor operations are examples of that) and irregular (Barnes-Hut as a very simplified example). To extract something like Barnes-Hut from the code with quadratic complexity, one have to make some interesting assumptions. E.g., one has to figure out something.

One may think that regular parallelism is a solved problem, but even nested loops are not always parallelizable without figuring out how to make them to work in parallel.

Then there comes Microsoft Word. Can we make a CPU where Microsoft Word gets executed with maximum IPC achievable? If you do not like MS Word, take a look at Linux Kernel and userland programs as a whole.