Talk:Vision processing unit

Untitled

If the AI accelerator article (currently in List of AI accelerators) gets 'consensus', maybe this article can be retargeted a little, avoid repeating material with that, which describes the broader class. I would not like to see them merged though. This one should describe vision in much more detail. Just like in 'graphics processing units' there were many classes over the decades. pure rasterisers, fixed function , scanline vs zbuffer, CPUs/DSPs with graphics functions, and eventually the current 'modern GPU' (which has more in common with a vector supercomputer than early graphics accelerators). I expect we'll see the same meandering evolutionary path with AI accelerators. — Preceding unsigned comment added by Fmadd (talk • contribs) 20:06, 17 June 2016 (UTC)

I don't really have a preference. If you aren't planning on radically expanding this article in the near future, it might make sense to redirect this page and its content to a subsection of "AI accelerator" until that subsection becomes large. Sizeofint (talk) 23:41, 17 June 2016 (UTC)

another option (if we get AI accelerator article) is that detail on 'VPU' could go in the Movidius (company) article, a layout that mirrors the Adapteva/Epiphany/Parallela(SBC) page. (Movidius/MyriadVPU/Fathom (USB stick)) Fmadd (talk) 23:47, 17 June 2016 (UTC)

Scope of definition of VPU

Hi, I am interested in helping with this article as I have some background in the topic and feel I can make a contribution. The first item I want to deal with is the scope of the definition: I feel that the narrow definition of visual processing as being restricted to AI and machine learning is too restrictive to deal with the actual processors that are out there. Case in point is myriad2 from Movidius, which has much wider functionality than machine vision since it includes functions that are normally regarded as computational photography and if you examine the specs of other devices in the space, you will see similar functionality.

It is true that there is a case to be made that the combination is temporary and that eventually the economics of the industry plus diverging tech requirements will at some point force a divergence but for now, at least, the two things are seen together in every case other than block IP.

I'm confident on my characterisation of the situation so I am seeking opinions on the best way to proceed, given that some attention will have to be given to hardware acceleration of computational photography and it could be appropriate at the outset to combine the two (leaving the potential to split them out at a later date). Pyjamaslug (talk) 06:43, 20 December 2016 (UTC)

You're welcome to take a stab at it. Preferably, content should be sourced from reliable sources. Fmadd may have some opinions on this, though it doesn't seem like zhe has been too active lately. Sizeofint (talk) 08:21, 20 December 2016 (UTC)

I'll draft something for review. Question: what is the accepted procedure? I saw on the compute project page a lot of references to putting stuff out for peer review but no hints on how to do that. Pyjamaslug (talk) 02:24, 21 December 2016 (UTC)

hi again, I'd be very interested to see any improvements you can make here Fmadd (talk) 03:10, 21 December 2016 (UTC)

If it seems controversial it might be worth posting it here first. Otherwise you can just edit the page. We'll be able to see the changes between revisions and can help edit it. If it doesn't seem like an immediate improvement to the article, an editor may revert the edit. Then a discussion should be held on the talk page. Sizeofint (talk) 06:46, 21 December 2016 (UTC)

Hi Pyjamaslug ; I hope you can re-work or extend these pages with your experience.

Let me elaborate what I was doing:

I was interested in having pages to define terms, including this ('VPU'), but I can see what happened: I subsequently wrote the 'AI accelerator' page to deal with many devices appearing for these kind of 'input-driven' workloads (movidius, TPU, truenorth);

I think that 'VPU' is really Movidius's own term, and the information might better go on the original 'Movidius' page? 'AI accelerator' would point at that (as an example). If you read the AI accelerator page, you will see I draw a parallel with the rise of Graphics Processing Units, when many overlapping devices used their own vendor specific terms. I personally have no experience of the VPU; I am merely commenting based on what I read in news reports. I am following AI and I have a background in gamedev. I find these processors interesting because similar units appeared in gamedev (e.g. CELL) in the past. I'm fascinated by how the space is unfolding.

For further reference, check out the page Physics processing unit for another time a vendor tried to make a term for a new class of device it was creating.

is there a 'category: accelerators' ? Fmadd (talk) 16:28, 22 December 2016 (UTC)