interesting talk r.e. general purpose machines (RISC-V)

Forum for anything not suitable for the other forums.

interesting talk r.e. general purpose machines (RISC-V)

Postby dobkeratops » Mon Mar 28, 2016 10:51 am

from the person pushing RISC-V ..

https://www.youtube.com/watch?v=VtJthbiiTBQ

I enjoyed this talk r.e. my thinking that I'd prefer a manycore RISC NOC (like epiphany) over dedicated outright NN hardware. (general purpose, just with a few commonly applicable specialised instructions like low precision dot-products or whatever).

He covers past attempts to build AI hardware (both LISP and neural nets) , compares to the current situation with CPUs + GPUs, and seems to argue that a general purpose machine with specialised units for recurring elements of common parts, but no specific algorithms 'burned into' hardware. Also says something about data movement being more important than computation (which seems inline with the epiphany philosophy).

He seems a big fan of vector machines (and notes a past attempt at a neural machine which he says is the same as the modern CPU+SLI GPU arrangement); but I'm sure there would be a sliding scale (N cores x M vector width). can the vector machines do low precision well? I think NN's (& vision) needs low precision (8bits for synapse/filter weights). I've always perceived packed-SIMD as being fine for this (if you have a dot-product instruction- 'horizontal add' ; pop count is just a 1-bit horizontal add).

He thinks GPUs will go away, absorbed into the CPU as an increasingly general vector engine (I guess intel KNL is the closest to this picture, with manycore+vector units, are there plans for similar RISC-V machines?)
dobkeratops
 
Posts: 189
Joined: Fri Jun 05, 2015 6:42 pm
Location: uk

Return to General Discussion

Who is online

Users browsing this forum: No registered users and 10 guests

cron