Floatplane Encoding Infra

[blakktux]

I saw a WAN show clip where Luke was griping about encoding WAN show content.  What’s your ingress pipeline like at Floatplane?  Are you running on a cloud provider?

 

You can parallelize video encoding by slicing the video into pieces then encoding, and stitch the results together after.

 

@Slick @LinusTech

[mariushm]

Maybe they can talk to AMD to get one of those cards for testing... but I think the big issue will be the API or whatever the cards will have... there needs to be updates to ffmpeg or other tools to take advantage of such cards, or have AMD give the source code for some basic command line encoder, otherwise who's gonna code some encoder from scratch.

 

I'd like to see at some point a card like this with some stupid simple api and interaction through ethernet (or have the driver make a server of sorts), and push video to encode though plain tcp or udp, and retrieve encoded video through equally simple requests. Then it would be stupid simple for anyone to use such cards.

For example, have a program create an encoding  by accessing a url of the built in web server on the card and you get a unique ID with the encoding job, then configure the created encoding job by sending a list of parameters you want to set, and then all the program would have to do is upload raw or encoded frames (in supported formats) to the card, and retrieve chunks of encoded content as it's done encoding.

[manikyath]

from what i've gathered, they <try to> spread the load as much as viable, with as wide spread internationally as viable. the priority being on "serving a video platform" for paying users over behind the scenes optimizations.

 

i think OVH was mentioned at some point, but i'm pretty sure nowhere near the entire party is on there.

 

as for parallellizing encoding.. i'm pretty sure they've talked about that in the past, but i assume it's a matter of:

- developer hours spent

- infrastructure requirement (more parallellizing means more cores, means more cost.. especially if they're sitting idle the rest of the time)

- prioritizing parallel processing of multiple videos being ingested (multiple creators / channels) before parallellizing the ingest of the WAN show (which.. is 3-4 hours of video recently..)

- slicing and stitching using resources as well.

- this whole thing needs to be vaguely profitable, and a huge influx of users doesnt immediately mean they *instantly* poop out more nodes.

 

also - a card like in this clip is not something you can just "plonk into a system" and use. it's most likely down to custom software to offload the workload to this card, in a similar way as it would be offloaded to a GPU.