Encoder/GPU for Plex in new server

Sawtaytoes · June 11, 2024, 9:03am

Problem

I just found out TrueCharts is ending support for TrueNAS Apps. This is where I’ve been putting Plex, Jellyfin, Pi-Hole, Tailscale, and a few other server apps. In the future, I might also run video processing on this box as well.

Requirements

The purpose of this box is to run Plex and any other containerized server software. The box won’t be doing much and won’t be using much RAM either. It also won’t have a ton of storage. A mirror of two 4TB drives should be sufficient. The rest will stay on my NAS and be accessed over the network.

What I wanna know is if there are any good 2U encoders or graphics cards I could use, or if should get a 4U case instead.

Server Hardware Options

I’m going to build a new server that runs Proxmox or TalosOS to host my server apps.

I thought about using a Raspberry Pi, but I think it wouldn’t be fast enough, and it’s got very limited RAM. Plus, I can’t easily add a GPU for Plex transcoding. It’s rare that I transcode, but HDR phone videos sometimes need it because they’re high-bitrate: 150Mb/s.

I have a bunch of consumer GPUs and even Epyc server CPUs I can use as well as RAM, I just need to see if I can find something that fits into 2U or if I need to go 4U.

The Case

I have a bunch of these 2U SuperMicro 216BE1C-R741JBOD cases that I think can take a regular motherboard:

If not, then I have to buy a new case anyway and would like to know some suggestions that’d fit my needs.

pkt1213 · June 11, 2024, 10:52am

A couple of GPU options would be Tesla P4, Intel A310, and Intel A380.

Sawtaytoes · June 12, 2024, 2:59am

That’s great! I looked them all up and yes, they exist in low-profile. I’m not sure about their capabilities though. How many transcode streams? What formats? etc.

Is that information easily available?

I’ve never heard of Nvidia Tesla. Is it older? Good to use? Or would the newer Intel cards be better because they have potential resale value? I’m not aiming to sell anything here though.

pkt1213 · June 12, 2024, 3:48pm

If I am still tracking correctly, NVIDIA is the only officially supported GPU by Plex (outside of intel igpu). I read a bunch of posts recently about the A310 and A380 working with no issues.

Tesla cards are datacenter cards. P4s don’t have active cooling but there are a lot of options based out there for 3D printing fan mounts. The P4 is basically a 75w 1080. Dual NVEC encoders. Lots of tutorals out there on getting PCIE pass through set up or even splitting the resources to different VMs. You can check the specs but they do everything except AV1 I think.

That said, I have a P4 which for some reason Plex isn’t using. I have passthrough configured. Plex sees it as an option for transcoding. But it isn’t using it and I haven’t had time to troubleshoot.

If you strictly want a low power card to do transcoding, the A310 seems like the best option. If you want to do more, the P4 or A380 make more sense IMO.

wertigon · June 12, 2024, 8:56pm

+1 for the A310 / A380 as a cheap-ish transcoding card. A310 is really all you need for 99% of transcoding needs, a friend runs it, and his card performs admirably

As for server CPU options, depends if you want low power, or not. EPYC is designed for efficiency, meaning it has a really good per-core perf/watt ratio but with that many cores, you can easily end up with 100W+ even in idle. After all, if you have a 32 core CPU and each core “only” draws 2W, that is close to 65W in idle, for CPU alone.

Meanwhile AM4 systems can get below 30W idle for a full system, and AM5 can dip below 45W. AM5 is great perf / watt wise and AM4 has a great power curve but is not as powerful as AM5. Asrock makes some AM4 and AM5 server motherboards, as do Supermicro and Gigabyte, if you want to go that route.

nubzzz · June 12, 2024, 11:29pm

+1 For A310 / A380. I just recently added an A380 into my plex machine and also added one into my kubernetes cluster for my own personal media app/AV1 support. The ASRock A380 that I picked up is half height/low profile as well so you wouldn’t have to bump up to a 4U chassis. That and it just runs off the 75W from PCI-E, no external power needed.

nx2l · June 12, 2024, 11:39pm

+1 For A310

I have one in my jellyfin system

Sawtaytoes · June 13, 2024, 7:49am

Yeah, I read that Intel Arc works well for Plex in Linux (native in Ubuntu). Sounds like a good option!

What’s the difference between the A310 and A380 for transcoding?

The Case

I’m gonna have good airflow since this is a rackmount server case, so I’d actually prefer a GPU that didn’t have separate consumer-style case fans. No big deal if they do though.

Another option: RTX 4060

One of the guys in my area suggested using a 4060. I think only the 4090 has 2 encoder chips though right? Or does this as well?

I’m curious if there’s really any difference between the transcoding performance of these cards. How many transcodes can you do at a given time?

How high of a bitrate can it transcode? 4K@60Hz with 10-bit HDR at 150Mb/s? That’s what my phone is recording and where I need to be able to transcode.

I know you can’t transcode with multiple GPUs in Plex, so I want it to be somewhat powerful even though I’ll very rarely need to transcode.

I also thought Nvidia GPUs had trouble running in Linux. Is that true?

Why Epyc?

This system isn’t just running Plex and Jellyfin, it’ll be running all the apps I had from TrueNAS. Plex and Jellyfin are the heavy hitters though.

I’ve already got the Epyc processors. I also have a spare AM4 5600X, but I’d rather use server gear for servers. It’ll be easier to rackmount, and I already know which cooling solution I’m gonna buy.

My plans

I’m debating having an 8-core Epyc vs a 16-core Epyc for this box (since I have both). My NAS has a 16-core right now, and I’m thinking of swapping it for an 8-core, putting a 16-core in this new app server, and use the other 16-core for a rackmount gaming rig I’m working on for a separate build.

Here’s the set of software I’ll be running based on my needs today:

Plex
Jellyfin (hardly ever gets used except when reviewing my YouTube videos since it does 2x)
Pi-Hole
Tailscale
Local IP camera video processing or accumulation. If I did this, I’d add a bunch of drives and probably wouldn’t use a 2U case, or I’d store the video data directly on my NAS.
Potentially moving Home Assistant off the Raspberry Pi 4 if it ends up being too slow. Some people said the Raspberry Pi isn’t powerful enough; although, mine has 8GB of RAM, and I’m not sure if they’re referring to older models.
Dual 10Gb links. I might pop a ConnectX-6 with dual 25Gb.

Does that 16-core Epyc sound reasonable for this build or would it be better for ZFS and TrueNAS?

nubzzz · June 13, 2024, 11:23pm

If it was me, I would put the 16-core Epyc in the app server and put the 8-core in your NAS. ZFS doesn’t really require a crazy amount of CPU and the 8 core should be plenty for it.

On the Arc front I am not sure what the difference between the two cards is. I can tell you that my A380 comfortably handles transcoding AV1 to 10Mbit H264 for about 10 users at a time.

pkt1213 · June 14, 2024, 7:22pm

I got my Tesla P4 working. It was operator error (kinda). The NVIDIA drivers weren’t installed. I installed them several times but they didn’t work.

I apparently don’t have any 4k movies, just 1080 from blueray. I have 3 being transcoded down now, 2 to 720 high and one to 420. My P4 is reporting 52-53C and 25-26W. With a single transcode it was 50C and 24-25W. This is with the faceplate off the GPU and a 40mm fan ziptied to the heat sink. 5.5GB of RAM.

I am running Ubuntu server with 4 Zen 2 cores assigned (3600) and 8GB of RAM. I am using under 5% of CPU and almost 100% of the RAM.

Sawtaytoes · June 19, 2024, 5:33pm

Thanks for the help and suggestions. I’m gonna try the “Gigabyte GeForce RTX 4060 OC Low Profile” for now, and if that has issues or problems, I’ll go for the “Sparkle A380”.

I don’t really want a gaming-style GPU in my server, but I’d rather try the top-end out for now. NVIDIA also made it so these cards can do up to 8-streams at once, and I only ever need 0; maybe 1.

Sawtaytoes · June 22, 2024, 6:37am

Now that I think about it, I’m running Plex/Jellyfin in a VM through Kubernetes.

Won’t that have an issue with NVIDIA? I thought they don’t allow SR-IOV? Or is that only when using the display outs? Should I have bought Intel instead?