Transcoding with 1050Ti hitting thermal limit

[hooraah 5]

So I finally pulled the 'broken' 1050ti out of my son's PC, got him a new card, and upgraded my Emby server with a 1050ti to do hardware transcoding.  It turns out that both of the small fans (its a low-profile card) were broken.  I thought about rigging up some kind of generic fan to the card, but then I thought that maybe if I could get by without a fan, I would have a nice silent option.  It has a fairly beefy heatsink on it for a low profile card.

Since I'm posting this from work, I don't have any screenshots to share, however, when running 4K transcodes, the temperature creeps up from 40C idle to between 85-95C.  At 90-95C (I don't remember what the number is) MSI afterburner logs the parameter "THERMAL LIMIT" going from 0 to 1, so it turns on some kind of thermal throttle to protect the card and the temp plateaus at 90-95C.

Here's the thing though - the GPU usage sits around 5%, and when the thermal throttle hits, I don't see the clock speed change, so I'm not exactly sure what the GPU is thermal throttling.  

Here's my question - can I continue running like this? 95C is high, but probably not catastrophic for the card. I could maybe try undervolting it or underclocking it, but I don't see that the clock speeds matter all that much.  I would prefer not to put a fan on it as I cannot find an off the shelf low profile fan that fits this card, and anything I come up with is going to be janky at best.  At some point, if I kill the card, I will probably just upgrade to an A310 from intel when the emby team releases AV1 transcoding.

[Q-Droid 989]

Are the broken fans still on the card or did you remove them? If you take off the fans and shroud then the bare heatsink might benefit passively from case airflow, though likely very little.

It'll work until it doesn't but it's only a 1050ti...

[hooraah 5]

41 minutes ago, Q-Droid said:

Are the broken fans still on the card or did you remove them? If you take off the fans and shroud then the bare heatsink might benefit passively from case airflow, though likely very little.

It'll work until it doesn't but it's only a 1050ti...

I completely took the fans and shroud off.  The fans that came on the card are integrated into the shroud, so if I wanted to put new fans on I'd probably have to either 3D print a whole new shroud or just zip tie something in place.  The system is a Dell Precision T5810 workstation, so all of the fan headers for case fans are proprietary Dell.  I could get adapters, but then at some point I'd probably rather just dump a $95 A310 in and call it a day.

[Neminem 1518]

I remember jerry-rigged a old GPU back in the early 2000.

Are there any heat sink left on the GPU.

There might be some groove where you can screw on some new fans.

[Neminem 1518]

Or try something like this

Universal Vga Cooler Dual 80mm grafikkort køleplade Dobbelt tre blæser Gpu Radiator Partner Ultra High Speed Stille | Fruugo DK

[RanmaCanada 494]

Honestly just dump an A310 or A380 into the machine.  The 1050Ti you could still sell for about $30 to someone who doesn't care if the fan is broken (I bought a 1060 with a 120mm fan zip tied to it for $25 CDN haha)

[rbjtech 5284]

A310/A380 is the obvious choice as it will destroy the 1050Ti in all aspects, but another option is to just redirect part of a case fan airflow over the heatsink - this will keep it under the thermal throttle which is all you want.    Janky as hell but a bit of cardboard cut from a cereal packet and some cable ties will work wonders ... 

[hooraah 5]

Just following up - it turns out that as the card hits its thermal limit, it IS dropping the clock speed (I was looking at the wrong graph earlier).  It eventually stabilizes at 90C and will sit there indefinitely until the transcoding request ends.  I was going to say that this seems to be sustainable, but then I got a complaint from a user that the transcodes were going slow.  Looking at logs, the server had 5 people on it at the time and I had exhausted the 3-decode limit of the 1050.  This is a new thing for my server, so I either need to do the driver hack or just get the intel card.

Now I just need to pick which intel card I should get.  I want to stick with something that runs off of just PCIE power with a low profile bracket so it can be moved to a slimmer system in the future, but on the other hand there have been nice price drops on some A750 cards that are more versatile in the future.  Not sure what to do...

[rbjtech 5284]

6 minutes ago, hooraah said:

Just following up - it turns out that as the card hits its thermal limit, it IS dropping the clock speed (I was looking at the wrong graph earlier).  It eventually stabilizes at 90C and will sit there indefinitely until the transcoding request ends.  I was going to say that this seems to be sustainable, but then I got a complaint from a user that the transcodes were going slow.  Looking at logs, the server had 5 people on it at the time and I had exhausted the 3-decode limit of the 1050.  This is a new thing for my server, so I either need to do the driver hack or just get the intel card.

Now I just need to pick which intel card I should get.  I want to stick with something that runs off of just PCIE power with a low profile bracket so it can be moved to a slimmer system in the future, but on the other hand there have been nice price drops on some A750 cards that are more versatile in the future.  Not sure what to do...

If it's purely for transcoding - then the only difference will be memory capacity which is only +2GB (8GB Total) vs 6GB for the A380 at 1/3rd of the price..     The huge bump in 3D performance for the A750 will not be used for transcoding.

[hooraah 5]

To help out anyone in the future, I thought I'd update this thread.  The 1050Ti ran for 6 months with no issue banging off of the thermal throttle limit.  I never really paid attention to it, I just let it do its transcoding thing in the background.

Last month a friend asked for any old GPU for his kids minecraft computer, so I zip-tied a fan on the 1050Ti and gave it to him and pulled the trigger on an Intel Arc A580.  The A580 was great, because it could transcode anything, but then the whole system went unstable.  It was constantly restarting and sometimes just hard-locking.  Running furmark would consistently trigger a system shutdown.  Despite the Dell Precision system having something like a 600 or 800W power supply, it did NOT like this card.  I took the A580 out, and the system went back to being stable, but now I had no transcoding at all because I gave the 1050Ti away.

Fortunately, I had an old Quadro K620 that at least gets me H264 hardware encoding and some H264 decoding, which helps.  So my production emby server is running that, keeping my clients happy, while I beta test another old system with the A580 in it that seems to be doing well.  Still, I want to do a lot of testing on it before moving it over for my family.

Lesson learned - I should have gotten the 310 or the 380 and been done with it.  Oh well.  Hopefully the 4770K / A580 combo will do well for the next couple of months to a year and I can build a new 14th or 15th gen system if the degradation issue gets worked out.

[matsonjr 6]

Did you use DDU when you swapped GPUs?

[RanmaCanada 494]

Does the Dell support resizeable BAR? Intel cards need that to function at their best.

[hooraah 5]

13 hours ago, RanmaCanada said:

Does the Dell support resizeable BAR? Intel cards need that to function at their best.

It doesn't, its way too old for that.  I'm not gaming on it (at least not on this machine), just transcoding is fine.  I put the A580 in another machine that also does not have resizeable BAR and it has been working well at 4K HDR 10bit transcodes.