I found reference to the issue on the same page you linked, and it lined up with what I was seeing during transcoding. The 1650 could only utilize about 30 or so percent of the NVDEC decoder per stream. You can do multiple lower quality streams, but it struggled with a single 4k bluray rip. There’s also the confusion that some newer 1650s use the Turing encoder, but the older models use pascal. Not that it matters here since the memory seems to be the limiting factor. Here’s reference to what I saw on that page.
VRAM Bandwidth
"Preliminary testing sees a single NVDEC job on a 128Bit GTX 1050 Ti 4GB (Pascal) unable to use more than 30% (112 FPS) of the NVDEC. Two streams hold about 50%, Three about 80% and more than four streams to reach 100% NVDEC saturation. A 256Bit GTX 970 4GB (Maxwell 2nd Gen) can hit 100% NVDEC saturation (376FPS) with a single stream.
The difference between the 128Bit, 192Bit and 256Bit Memory bandwidth needs further testing. It looks like a 128Bit memory bus will not cause performance issues with multiple transcodes but will see Plex offline “Sync” jobs only able to use 30% of the NVDEC chip."
If you’re using a 2U rack mount server, you may look into a PCIe riser to get around your half height limitation. As far as I know 1660s are only available in full height, but I believe some come with single slot layouts. I’m a server engineer and at work we have Dell PowerEdge R640s that are 1U but have full height single slot quadros in them using risers. Depends on the model of course if that will be an option for you.
It appears the chart is wrong based on my testing of the 1650. I had the newer gddr6 version, but it still had the same issues you’re describing due to the 128bit memory interface. I can confirm with 100% certainty that a 1660 ti is fully capable of transcoding multiple 4k streams, and I’m almost positive a regular 1660 could do the same because they share the same 192bit interface and hardware decoder. I know pricing and supply are crazy right now, but at least you know what you need. Sorry I don’t have better news for you.
as much as ide like to believe memory is the limit here.. it really makes no sense.
128 GB/s is what the gddr5 version has. 192 is what the gddr6 has.
“The GTX 1650 reportedly has 896 CUDA cores and 4GB of GDDR5 memory . It’s expected to come with a 1,486MHz base clock and 1,665MHz boost clock. The memory is clocked at 2,000MHz (8,000MHz effective) across a 128-bit memory interface, which works out to a memory bandwidth of 128 GB/s”
128GB/s that more than enough
even if it wasnt then the 1650 gddr6 version has a 50% boost over the reg 1650.
"The memory is the most obvious upgrade. The original GTX 1650 leverages 8 Gbps GDDR5 memory to offer 128 GBps of memory bandwidth. The latest revision flexes 12 GBps GDDR6 memory and is good for up to 192 GBps of memory bandwidth. The upgrade is equivalent to a 50% increase. "
288 GB/s
The GeForce GTX 1660 Ti also features 6GB of GDDR6 memory operating at 1,500MHz (12,000MHz effective) across a 192-bit memory interface resulting in a memory bandwidth of 288 GB/s
so if the 1650 cant do one 4k with 192GB/s how can the 1660 do 4 with 288GB/s?
the math doesnt add up.
seems more like a plex bug with the 1650 or driver issue
Sun Feb 21 10:44:46 2021
±----------------------------------------------------------------------------+
| NVIDIA-SMI 460.39 Driver Version: 460.39 CUDA Version: 11.2 |
|-------------------------------±---------------------±---------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 GeForce GT 1030 Off | 00000000:08:00.0 Off | N/A |
| 40% 42C P0 N/A / 30W | 1322MiB / 2000MiB | 16% Default |
| | | N/A |
±------------------------------±---------------------±---------------------+
±----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 1295 G /usr/lib/xorg/Xorg 55MiB |
| 0 N/A N/A 1366 G /usr/bin/gnome-shell 6MiB |
| 0 N/A N/A 4308 C …diaserver/Plex Transcoder 1257MiB |
±----------------------------------------------------------------------------+
I hope the devs will fix this soon.
BTW, my e5 can handle 1.5x a 4k transcode on its own, so that is not the issue.
Is tone mapping single core?
Edit. Just checked. Tone mapping enabled using CPU transcoding does work, but not with 7.1 audio
GT 1030 combined with cpu encoding + tone mapping does buffer, even with 2.0 audio
The chart specifically mentions the memory interface of 128bit being the issue, not total bandwidth. The quadro p2000 has a 160bit interface. Perhaps the memory pipeline is not wide enough to fit a single uncompressed 4k BD stream? I’m not sure of the technicalities behind the memory interface issue, but it’s clear overall bandwidth is not the problem. I’d be curious to know the bitrate of your files to determine if it’s an issue with resolution in general. The files I tested were 4k bluray rips.
Another test you could try is to use a different program such as handbrake to encode your files. If you can encode the same 4k files at more than 1x speed it could prove there’s an issue with the Plex transcoder and the 1650/128bit cards or drivers.
i may have something wrong with my plex installation. now when i have hardware selected to use gpu with tone mapping off i cant even transcode. get this error An unknown error occurred (4294967279) when using web interface or windows plex player. fire stick at least tried to play get the every 15 sec freeze. this is with a blue ray rip and i have a 16mbps 4k as well. both do the same thing
i can direct play everything fine. if i turn off hardware support my cpu ramps up just fine and can transcode even 68mbps blue ray with 70% cpu usage.
im going to dd remove drivers today and install 2 revisions newer and see what happens
at this point if it doesnt work i may just add a 2nd e5-2697 v2 for 120$ and just removed the gpu. wanted the power savings from the gpu but its been a pain. at least i know 1 cpu can do 1 4k transcode at least.
I have comparabe issues. CPU transcoding used to work, even with hdr tone mapping. Now it does only work without tone mapping.
Gpu transcoding issues aswell. I thought of adding a gt1030 to my setup for the decode part, since cpu only can only handle 1 transcode. So, I thought with gt1030 I might do 2-3 transcodes of 4k content.
These issues seem to happen to emby aswell.
I have this weird system that allows RAM memory to be shared with the nvidia dedicated gpu. Maybe issues with that?
Anyway, issues with transcoding made me sell one of my gt1050’s in the past, because transcoding didnt work all together, which seems the same issues you are having. That was back in 2018…
