I’ve got an old AMD Phenom II system in a 2U rackmount case that I want to run Jellyfin on. I need to add a graphics card so I can hook a monitor to it for troubleshooting etc., so I figure while I’m at it I ought to find something I can pass through for hardware-accelerated transcoding.
I use Linux and I prefer open-source drivers, so I’ve initially limited my consideration to AMD. Based on this page in the Jellyfin documentation, it sounds like I need GCN 5 (Vega) or newer for acceleration on Linux with the open-source driver, or GCN 1 or newer for acceleration on Windows with a closed source driver (not sure if that includes Windows in a VM on Proxmox?).
Needing a low-profile card really limits my options. Looking at Ebay etc., the ones with decent availability include:
- FirePro W2100 (GCN 1) (~$12)
- Radeon R5 430 (GCN 1) (~$12)
- Radeon R7 450 (GCN 1) (~$30)
- Radeon Pro WX 2100 (GCN 4) (~$35)
- Radeon RX 550 (GCN 4) (~$45)
- Radeon Pro WX 4100 (GCN 4) (~$70)
- Radeon RX 6400 (RDNA 2) (~$130)
Then I noticed a caution on the Jellyfin documentation page:
Most AMD dGPUs come with video encoders but be careful with certain models - RX 6400/6500 series don’t have video encoders.
So that eliminates the only GCN 5+ low-profile card that AMD has apparently ever made, which means there is no such thing as a low-profile AMD card that both supports VA-API and has a video encoder? Is that really true that AMD went the five years between the WX 2100 in 2017 and the RX 6400 in 2022 without releasing a single low-profile card suitable for Jellyfin??
As for Nvidia, I’m starting from scratch because I haven’t bought an Nvidia card in probably 20+ years. According to the Nvidia Jellyfin documentation, apparently transcoding requires the proprietary driver and a card that’s Maxwell (1st gen) or better, HEVC requires Maxwell (2nd gen) or Pascal depending on bit depth, and AV1 requires really new stuff that’s probably out of my budget.
Nvidia low-profile cards AFAIK include:
- Quadro K620 (~$25)
- Quadro K1200 (~$40)
- Quadro P400 (~$50)
- Quadro P600 (~$60)
- GT 1030 (~$70)
- GTX 750 (~$100)
- GTX 1050 Ti (~$100)
- Quadro P1000 (~$150)
- T400 (~$160)
- GTX 1650 (~$200)
- RTX A2000 (~$300)
TLDR: so it seems like my options are basically to get an AMD FirePro W2100 or Radeon R5 430 for $12 and resort to running Jellyfin in a Windows VM, to get a Quadro K620 for $25 and resort to using the proprietary Nvidia driver, spend way more than it’s worth for anything that can do HEVC (let alone AV1), or scrap the whole CPU/mobo/RAM and upgrade to something with modern integrated graphics?
It’s not actually available (yet), but why not consider a low profile A380 like this
Relatively cheap, open source drivers, and hardware encoding up to and including AV1.
Edit: oh, and supported on jellyfin. I have mine running on truenas in a fedora 38 VM by passing the GPU through(Truenas scale is running kernel 5.15, so doesn’t have native support).
That’d be great, except for the whole “not actually available yet” part!
(Well, that and the price: since I don’t actually care about 3D performance, I was hoping to get something old that only cost two digits.)
I guess I could get a $12 AMD card now (just for hooking up a monitor, not necessarily for Jellyfin) and then get an A380 LP once it’s actually available. It really is a shame AMD broke the RX 6400 for this use-case, for no apparent reason…
Incidentally, this article about that card claims:
What other Intel low-profile graphics cards are there?!
Well to be fair, the A380 runs circles around every cars you mentioned in encoding.
The A380 at launch beat the 3090 at encoding H.265 and h.264 and is the single cheapest AV1 transcoding card.
It has insane value. You wouldn’t buy it for 3D gaming because it has mediocre 3D performance for its class and the ARC cards in general absolutely suck on linux for gaming.
I believe it’s talking about This MSI card, which seems to be prebuilt only? Or maybe the Sparkle Genie A380, which, again, doesn’t seem to be available yet.
Are you able to pass the prime GPU to a VM now in Scale? When I tried it with a 750Ti, I had to reserve one GPU to Truenas Scale itself, which was not convenient. I haven’t tested lately since the latest Scale upgrade.
No, you still need display output of some type