What a weird article. Maybe this is too intellectual for me, but it makes almost no sense. It's like they wanted to write about DPUs without ever saying "DPU" and this was the result.
What a weird article. Maybe this is too intellectual for me, but it makes almost no sense. It's like they wanted to write about DPUs without ever saying "DPU" and this was the result.
At first I wondered if the article was poorly translated to English. It seems like they were using the term Operating System to mean something other than what I am familiar with. They are either...
At first I wondered if the article was poorly translated to English. It seems like they were using the term Operating System to mean something other than what I am familiar with.
They are either talking about something so extremely specialized to the world of ML development that none of us plebeians can understand it, or they wrote all of this without really understanding what they were talking about. Or it was written by an AI.
I too may be missing something, but I fail to see how the new proposed paradigm is significantly different. Instead of a GPU you have an AI-PU. Instead of on-die instruction sets you have seperate...
I too may be missing something, but I fail to see how the new proposed paradigm is significantly different.
Instead of a GPU you have an AI-PU. Instead of on-die instruction sets you have seperate chips. At the end of the day it still needs exposed as a capability to userspace.
I don't see how any of this is any different from having PCI-E lanes connecting different coprocessors.
I feel like even if it moves in this direction, it'll eventually shift back when someone remembers that it was a lot easier to do stuff when you had a single coherent OS layered on top of the hardware OS (what we call the BIOS now).
What a weird article. Maybe this is too intellectual for me, but it makes almost no sense. It's like they wanted to write about DPUs without ever saying "DPU" and this was the result.
At first I wondered if the article was poorly translated to English. It seems like they were using the term Operating System to mean something other than what I am familiar with.
They are either talking about something so extremely specialized to the world of ML development that none of us plebeians can understand it, or they wrote all of this without really understanding what they were talking about. Or it was written by an AI.
I too may be missing something, but I fail to see how the new proposed paradigm is significantly different.
Instead of a GPU you have an AI-PU. Instead of on-die instruction sets you have seperate chips. At the end of the day it still needs exposed as a capability to userspace.
I don't see how any of this is any different from having PCI-E lanes connecting different coprocessors.
I feel like even if it moves in this direction, it'll eventually shift back when someone remembers that it was a lot easier to do stuff when you had a single coherent OS layered on top of the hardware OS (what we call the BIOS now).