The Kiwi Farms Media Processing Server

Null said:
want to write a media processor that converts user uploads into playlists.
Click to expand...
I looked at this a while ago for the BMJ VODs: https://github.com/kaltura/nginx-vod-module

It can repackage MP4 files into HLS and DASH on the fly, no re-encoding needed. Supports caching metadata, remote file systems, captions and clipping. Seems pretty comprehensive, I just haven't used it since my site is all just directory indexes and I can't be bothered making a real website that can serve a proper media player.
 
Null said:
ok chyat how about this

1x https://www.ebay.com/itm/127566125829
4x https://www.ebay.com/itm/297951440555

I think someone offered me a bunch of WD Enterprise 12TBs recently, I need to go find that email. It came at a really bad timing because I had already bought 16TBs for the server but it would actually be SUPER DUPER PERFECT for this
Click to expand...

The Supermicro's are Skylake era so they run at PCIe 3.0 speeds while the Arc Pro B60 runs at PCIe 5.0 x8, so you would have a small performance penalty in the amount streams you could transcode at once due to PCIe bandwidth limitations.
 
Last edited:
Drama Fan said:
This would be literally a dream come true for me and every lolcow's worst nightmare. I constantly want to reference random videos and can't because they're buried in several thousand page threads with no obvious search terms in the post.
Click to expand...
Off topic but are you familiar with Bookmarks? Not as useful as having every vidya transcribed, but you can at least tag things you want to come back to:

1770488351060.png

1770488310461.png
 
The best nvidia server card is surprisingly affordable
Great for AI and media processing
Buy the good nvidia server card and cut a hole in the server chassis for it. Plus 96GB of vram.

If you are focused on purely media encoding
The ARC cards are good, but they usually only have 16GB of ram, maybe the B60Pro?
Arc Pro B60, it requests qoutes though. Also size may not be right. But it has 48GB of VRAM shared with dual gpus.

What chassis is this server in? What RU? What power supply? It would be way easier if we knew the size/specs of this server.
 
JohnnyG said:
Off topic but are you familiar with Bookmarks? Not as useful as having every vidya transcribed, but you can at least tag things you want to come back to:

View attachment 8530350

View attachment 8530346
Click to expand...
yeah, i use those, but not religiously enough. My favorite one is this Philly guy talking about how they were justified to have a stadium full of people throwing batteries at santa
Tiggletown

Post in thread 'The NFL Thread'

Being a Philly sports fan I was an adult when I found out other teams didn't act like total Niggers when their teams win or lost.

I'm honestly saddened for the rest of you, I always had so much fun taking the Regional Rail down to city hall and destroying stuff and partying. It's like a George Floyd Riot for white people with jobs, we just flip cop cars and climb poles we don't loot shit.

I remember emptying my CC into the air when the Eagles won the super bowl like it was an ISIS wedding.

When people find out i'm an Eagles fan they say "Oh you threw shit at Santa" and I have to remind...
Reactions:Like x 11 Winner x 5 Informative x 2 Feels x 1 Thunk-Provoking x 3
1770489811390.png
 
I hate that graphics cards are now as expensive as used cars. Im going to cry when my lovely tower who has been running for a decade decides to cark it. State of things in this bloody world, good luck null!!
 
babaisyou said:
The Supermicro's are Skylake era so they run at PCIe 3.0 speeds while the Arc Pro B60 runs at PCIe 5.0 x8, so you would have a small performance penalty in the amount streams you could transcode at once due to PCIe bandwidth limitations.
Click to expand...
I really don't think we need 32gbps bus.
 
Just a heads up if your planning on using it for AI stuff, getting AI inference shit to work on non-Nvidia GPU's has been a pain in the fucking ass in my experience. But then you gotta deal with Nvidia's drivers and hope they work with whatever kernel you are using. Pick your poison I guess.

I'm seconding peoples opinions on using Intel GPU's for media transcoding, Intel has been in that space for over a decade, and their GPU's are comapratively cheap. And IMO they are more likely than AMD to get better AI support compared to Nvidia in the future.
 
Null said:
ok chyat how about this

1x https://www.ebay.com/itm/127566125829
4x https://www.ebay.com/itm/297951440555

I think someone offered me a bunch of WD Enterprise 12TBs recently, I need to go find that email. It came at a really bad timing because I had already bought 16TBs for the server but it would actually be SUPER DUPER PERFECT for this
Click to expand...
I like this idea. A good rack-mounted server with as many good ARC GPU's as you can fit in it seems like it would be a good jack of all trades for your use cases.
 
Here's my economic and opportunity cost argument since I know very little about this hardware case itself and more about business budgets and logistics around similar hardware choices from a super microniche R&D doillar-crunched perspective.

I recommend whichever system has the most redundant parallelism, is capable for OTHER use cases if needed in an emergency, least proprietary driver and software, and can still function with more individual hardware failures or can have similar capabilities on the CPU side and on GPU/encoding hardware/expansion card options side as may be needed. High-dollar corpo structures encourage single-item specific-use high-performance hardware but highly generic infrastructure. The items which do the work need to be very efficient to replace into any global infrastructure which can accomodate them (cough cheapest server farm on Earth with main power rails that look like square wave seizures on an Oscilloscope cough) and have very high performance capacities per "fixed cost operation" i.e. labor and space lease for budgetary and time reasons. Adding on 10 units that do 10 units of work is cheaper than 100 units that do 10 units of work: reduced complexity, faster to swap in and out, lower fixed costs, and therefore has less business risk IF YOU HAVE CASH TO EAT HIGH DOLLAR FAILURES! The theoretical fun things one can do with a single powerful GPU must be balanced with the cold hard reality of a dollar-limited business like this forum: you're going to want the most "I have had 2/3 break but I can still make the last 1 work because I need 6 months for replacements and the on-site install tickets are 3 weeks behind schedule and my other hardware broke so I had to use this to do something temporarily while that server is fixed" system. A single high-dollar item failing will be crippling. Absent the wisdom of a hardware specialist saying otherwise, a budget system with very robust IO expansion opportunities and clusters of lower-dollar cards would be my own personal choice if I were given the budget to approve and two choices.
 
Null said:
I don't think it's possible to build a server that does both media and AI without just getting big boy NVIDIAs, spending a ton of money, and plugging it into a 250W outlet.
Click to expand...
You're looking for a cross between an ASIC and GPGU. Sophgo maybe... In this area Intel/AMD should beat the crap out of Nvdia since they own ASIC subsidaries for a while now.

Niche card but AMD (technically Xlinix, however that's spelt) has a specific ASIC card for this, they call it a "Media stream accelerator". 22 TOPS， 16GB vram 40W TDP, 1.5k per card. Sad part is no cuda.

Edit: 1.5k, not 21.5k per card. Oops
 
Last edited:
Null said:
$6000 is honestly too much for our workload if I managed to get something like NETINT Quadra Cards but it makes me wonder if we could meme together something that can do both media + ai inference.
Click to expand...
I'm not experienced enough in this area to offer a hard opinion. But generally, I'd say focus on one thing and get something solid built for that. You're already getting complex with the equipment whichever goal you have, the complexity just scales up when you try to do both at once.

If you build out a media processor, it runs great for a few months, and you realize you have spare capacity for AI, then you could start planning something for that. If you build for both then find an unexpectedly high load from one, you'll just get frustrated trying to troubleshoot and balance multiple issues.
 
A piece of hardware that I'd like to bring up for media/stream transcode optimized workloads is Intel's Flex 140 from their Flex series data center GPU lineup. Although it's a bit older model from a few years ago, the 140 is half-height, it can be used for some lighter AI/deep learning inference tasks, and there seems to be a fair bunch of them that you can find on eBay and up for resale elsewhere.
 
Last edited:
You must log in or register to reply here.
Back
Top Bottom