The most impressed I’ve been with hardware encoding and decoding is with the built in graphics on my little NUC.

I’m using a NUC10i5FNH which was only barely able to transcode one vaguely decent bitrate stream in software. It looked like passing the hardware transcoding through to a VM was too messy for me so I decided to reinstall linux straight on the hardware.

The hardware encoding and decoding performance was absolutely amazing. I must have opened up about 20 jellyfin windows that were transcoding before I gave up trying and called it good enough. I only really need about 4 maximum.

The graphics on the 10th generation NUC’s is the same sort of thing that is on the 9th gen and 10th gen desktop cpu’s, so if you have and intel cpu with onboard graphics give it a try.

It’s way less trouble than the last time I built a similar setup with NVidia. I haven’t tried a Radeon card yet, but the jellyfin docs are a bit more negative about AMD.

Yep, any time you have a traffic cap or bill for traffic you’ve got to have data to back up what you are billing for.

More recently CDN’s ( and widespread SSL adoption ) have made it a whole lot less obvious what sites the user is going to. I suspect that nice clearcut list of porn sites from 2007 would just look like some cloudflare, akamai and google these days.

There’s no way of knowing what happened there.

But back in the mid to late 2000’s we had a whole bunch of residential internet customers and every so often one would blow their traffic cap by a bunch and would ring up and say “Your billing system is wrong!”.

Then whoever could be bothered in the office would do some modest analysis on their netflow data and come up with something like "18% of your traffic this month was redtube.com, 33% was pornhub.com and 9% was xhamster.com.

We never knew if whoever was on the phone was the raging porn addict or it was one of their associates. Either way they would say “Oh well, I guess we will never know then. Thanks for your help. Bye.”. Followed by them quietly paying the bill.

Haha, 144p @ 60hz is fricking hilarious.

Reminds me of seeing completely rubbish resolution real player videos embedded in websites back in the late 90s and me thinking, “Well that isn’t ever going to take off”.

The situation is mostly reversed on Linux. Nvidia has fewer features, more bugs and stuff that plain won’t work at all. Even onboard intel graphics is going to be less buggy than a pretty expensive Nvidia card.

I mention that because language model work is pretty niche and so is Linux ( maybe similar sized niches? ).

Please drink a verification can to continue.

The samsung TV that I bought for my son had this annoying overlay thing that pops up when you turn it on that shows all the different inputs and nags about various things it thinks are wrong with the world. It is plugged into an Nvidia shield that we do most things on, but you can’t use the shield until the overlay calms the fuck down and disappears.

It’d be great if you could just have the thing turn on and display an input like our older TVs do.

I swear allegiance to the only one true storage vendor, Micropolis. The Micropolis 1323A being the embodiment of perfection in storage basked in the glow of the only holy storage interconnect, MFM.

I wait patiently for the return of Micropolis so that I may serve as their humble servant.

Your bridge isn’t bridging properly. If Router B is sending a destination unreachable then the packets are being handled on it further up the stack at layer 3 by some sort of routing component rather than by a layer 2 bridging one.

Steam can do pretty well filling a tail circuit, probably better on average. But a torrent of a large file with a ton of peers when your client has the port forward back into the client absolutely puts more pressure on a tail circuit. More flows makes the shaping work harder.

Sometimes we see an outlier in our reporting and it’s not obvious if a customer has a torrent or a DDoS directed at them for the first few minutes.

No, if two 300 megabit tails are shaped correctly, a third user shouldn’t notice that the 1G backhaul has got a bunch of use going on.

If you do, there’s something wrong or you aren’t really getting the 1G for some reason. Not generally a concern in a carrier platform.

I build ISP and private data networks for a living.

A contention ratio for residential circuits of 3 to 1 isn’t bad at all. You’d have to get pretty unlucky with your neighbors being raging pirates to be able to tell that was contended at all. Any data cap should scare the worst of the pirates away, so you probably won’t be in that situation.

If you can feel the circuit getting worse at different times of the day then the effective contention ( taking into account further upstream ) is probably more like 30 to 1 than 3 to 1.