Re:Leveraging RISC-V for AI and Machine Learning
by CalChris on Saturday December 16, 2017 @04:26PM
(#15942219)
I wonder how Esperanto compares to Green Arrays [1] or REX [2]. Obviously, these are very different approaches and GA doesn't even support float. But REX does.
Re:Leveraging RISC-V for AI and Machine Learning
by baobrien on Sunday December 17, 2017 @08:40AM
(#15945717)
I'm not sure about REX, but Esperanto is pretty far away from greenarrays. From what I got out of their talk at the latest Risc-V workshop[3] (video[1], slides[2]), they're going for the big high-performance end of the spectrum. The chip they were talking about is supposed to have 16 64-bit OoO Risc-V cores and 4096 RV64GC cores with vector units connected by some sort of network on chip.
Re:Leveraging RISC-V for AI and Machine Learning
by jabl on Sunday December 17, 2017 @05:49AM
(#15945215)
Don't know much about Green Arrays. It seems slightly strange in that most tasks that can make use of massive data parallelism are FP-heavy (numerical simulations, ML/AI).
Re:Leveraging RISC-V for AI and Machine Learning
by godelmachine on Sunday December 17, 2017 @01:12AM
(#15944459)
Is Rex Computing still in business? Just yesterday I was thinking about the founder who had won the Theil prize. I don't think their website has any updates.
For those who follow RISC-V, this doesn't add any new information that wasn't already made public during the latest RISC-V conference. Proceedings here: https://riscv.org/2017/12/7th-risc-v-workshop-proceedings/
http://www.greenarraychips.com/home/products/
http://rexcomputing.com/
[1] https://www.youtube.com/watch?v=f-b4QOzMyfU
[2] https://content.riscv.org/wp-content/uploads/2017/12/Tue1136...
[3] https://riscv.org/2017/12/7th-risc-v-workshop-proceedings/
As for REX, I don't really know what to think about it. Here's a recent thread on RWT: https://www.realworldtech.com/forum/?threadid=172976&curpost... . People there seem quite skeptical about the whole thing.
1) Asynchronous logic. 2) Stack machine design. 3) Extremely limited memory per cell. 4) 18-bit words. 5) No MMU.
Chuck Moore gave a nice presentation on applications and his rationale here: https://www.infoq.com/presentations/power-144-chip
I thought it was pretty neat how in some of his design docs, a column of async cores are used as a memory controller.
It'd interesting to explore how Arduino and, later, the RPi impacted their respective architectures' presence in the embedded space.
https://www.youtube.com/watch?v=f-b4QOzMyfU