Transcoding 4k->4k with HDR tone mapping and nvidia hardware accel crashes server. power off

Couple things on specs… the 1660 comes up as pulling up to 120W… which is a huge chunk of your power budget right off the bat, and rings alarm bells in my head.

When I plug your cpu, ram, gpu into random internet calculator (even omitting storage, fans, etc) coming up to 220w load, recommending 270w…!

Also, PC power supplies have to supply a lot of different voltages, and they have different limits. 3.3v, 5v, 12v, and the neg polarities.
Out of the total 200w max for entire supply, it might can only supply 120-140w max on the 12v rail. While the combined power from 3.3 and 5v rails make up the remaining of the 200w.

So then you have a 20w or so cpu, 120w gpu, mainboard parts and storage parts all competing for that constrained slice of 12v power rail.

1 Like

The amount of power in the design spec and the amount of power actually used are totally different. And I accounted for that when putting the system together (before HDR tone mapping was a thing)

Even though the design spec is 120W, that’s only under 100% load running graphics or computational loads. The NVENC encoding used by Plex is very low power hardware. Previously when the worst we’d see was a 4k to 1080 transcode without Tone mapping, the GPU only used like 30-40w, well within the specs of the PSU. If I could have bought a lower power GPU with at least 6GB of ram and a Turing encoder, I would have. maybe with Ampere Nvidia or board partners will finally come out with a 75W GPU that has a good encoder and a good amount of vram.

But even with the tone mapping, we’re only up to 80w still should be in spec of a 200W PSU since the cpu is only using 10-20W. But it still remains to be seen if it’s just a faulty PSU or not.

You can’t see transient load and a power supply for sustained average isn’t adequate.

Average utilization is good for calculating power bills and heat output. :slight_smile:

If your system draws 1000w for only 100ms, and is then idle for 900ms, thats 100w average, but requires a 1000w supply.

There’s a lot of engineering slack built into power supplies and motherboards, but the start of a transcoding job is going to spin up every power-hungry part of the system, even if it calms down quickly. CPU, memory, busses, GPU, storage/network.

Max loads are good for accounting for transient loads, burst that happen when things start up. Generating buffers.
Transient, instantaneous loads can still trip OCP circuits and other issues.

Sadly, how much engineering slack built into pc power supplies is not a given. Many have been proven to not even meet their advertised design specs.

I was watching transient loads. The values I listed were the maximum instantaneous loads that I observed. Averages were 10-15W less.

It’s still well within spec.

swapped out the PSU with my spare (same exact model, Supermicro PWS-203-1H) and the problem has gone away, so I was right that the system isn’t pulling too much power for this PSU spec when the PSU isnt defective.

The PSU only has a single 12V rail spec’d for 16A, so thats 192W that is should be able to handle.on 12V alone, with 50W available to 5V and 3.3V.

Nice!

I’m suspicious that software is accurately reporting transients, but I’m also a big fan of you understanding your system and that you were right and that it’s working again now.

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.