-
SOLVED: Electrolytic SMD caps getting too hot to touch
Posted by
daxliniere
on 07 Aug, 2021 14:37
-
Hey gang,
I have tested two different brands of PCI-e riser boards with an audio DSP card containing 8x SHARC 21469 ICs. Both of the risers seem to work perfectly, but their capacitors get too hot to touch within about a minute.
The max current draw for that card is allegedly 8W, which seems to check out reading the Analog Devices datasheets. These are the kinds of cards that are used in cryptocurrency mining, so they normally have a whopping big GPU card plugged into them, not a little 8W DSP card. The risers have a 12V ATX connector (6-pin) for power input and send the data signals across a USB3 cable (not USB3 protocol, of course) to a dummy card that plugs into the motherboard's PCI-e slot. One of the riser cards I've tested has 4 electro caps the other has 8 caps (4x2 in parallel).
My questions are:
1) Why would this be happening?
2) What could I do to mitigate it? I'm happy to change or add caps.
Thanks in advance to anyone who can shed some light on this strangeness.
All the best,
Dax.
-
#1 Reply
Posted by
Siwastaja
on 07 Aug, 2021 14:47
-
DC/DC converter poorly designed.
By the looks of it, uses some ancient 1990's switcher technology with ridiculously small fsw, typical to Chinese ebay crap for whatever reason.
Modern switchers use higher frequency, allowing smaller capacitance and ceramic cap only solutions (with possibly electrolytics for input damping) and do not suffer from this.
You could try to replace those electrolytics with polymer types, or maybe just solder some largish (a few uF) SMD MLCCs in parallel with the existing caps. Ringing and oscillation is a risk when not knowing the existing design, though.
Adding cooling, or accepting short lifetime are always options as well.
-
#2 Reply
Posted by
NiHaoMike
on 07 Aug, 2021 14:55
-
Not much on a modern GPU runs on 3.3V. Perhaps the board is not designed to source much at 3.3V?
-
#3 Reply
Posted by
daxliniere
on 07 Aug, 2021 15:11
-
Thank you both for your replies.
I did a bit more poking around and the two caps that get the hottest are the cap that's in parallel with the 12V supply (which goes to the PCIe slot) and the cap that's in parallel with the 3.3v regulator's output.
As far as I can tell, it is sourcing all of its required power from the 12V ATX connector and regulating that down to provide the 3.3v, as opposed to sourcing that from the computer's PSU, which is a high quality BeQuiet! Straight Power 550W.
The board is quite minimalist. Protection diode, SMD fuse, 4 electro caps (270uF 16V), 3 amp DC-DC buck converter (Lite-On LSP5503L,
datasheet from 2014), 3.3v 5A regulator (
1084-33) and a handful of passive components (LCR).
-
#4 Reply
Posted by
daxliniere
on 07 Aug, 2021 15:40
-
Seems like the 3.3V regulator they chose has a maximum input voltage of 7-8V (depending on manufacturer), so I guess that's why they used the buck converter ahead of it.
(Why not just choose a LDO regulator that happily accepts 12v??)
I thought they might be sourcing 5V DC for the PCI-e card, but PCI-e doesn't require 5V.
-
#5 Reply
Posted by
Siwastaja
on 07 Aug, 2021 15:57
-
In a buck converter, input cap is stressed most. It's unsurprising the 12V capacitors are heating up.
If they are too high ESR, they can heat up even if the converter is supplying just a watt or two of output power. If output current is say mere 500mA at 3.3V and Iripple=40%, the output capacitors see 200mA of triangular ripple current but the 12V input caps see square wave 500mA ripple current at the f_sw (say 200kHz). This is over the 100kHz ripple current rating of even many such small size low ESR electrolytics, and if they are general purpose "normal ESR" ones even more so. That's why you need a better capacitor type, or if electrolytics are to be used, just larger/more of them to offer lower ESR (with the side effect of having even more unnecessary but not harmful capacitance).
-
#6 Reply
Posted by
daxliniere
on 07 Aug, 2021 17:57
-
Thanks again, Siwastaja. I will get some low ESR caps and solder them in parallel to see if that helps the situation and report back.
Cheers!
Dax.
-
#7 Reply
Posted by
daxliniere
on 07 Aug, 2021 17:58
-
While I think of it, would there be any value in putting a low value cap (ceramic? MKT?) across these overheating caps? Would that soak up some of the high frequency ripple?
If so, any suggestions for a value? 0.1uF maybe?
-
#8 Reply
Posted by
tooki
on 08 Aug, 2021 00:14
-
Stuff like this makes me feel two things:
1. Pride that my implementations of the designs TI’s webench tool spit out have worked very well, including the ones where I couldn’t copy their component suggestions and PCB layout 1:1.
2. Incredulity that the people who design these products aren’t smart enough to use tools like Webench, which literally lets a DC-DC converter novice like me produce designs that work well. Even if they ultimately use a Chinese clone part instead of the original, it’ll perform better than the OP’s board.
-
#9 Reply
Posted by
NiHaoMike
on 08 Aug, 2021 01:02
-
It looks like it works so poorly because they're using a linear regulator for final regulation when there shouldn't be any need for one. My guess is that the designer couldn't properly design a switcher and ended up making one that's far too noisy to work, but raising the voltage some and adding a linear regulator worked around it. With the mining market moving so fast, the next step was to just mass produce it and not bother with redoing it "right".
-
#10 Reply
Posted by
Siwastaja
on 08 Aug, 2021 06:40
-
While I think of it, would there be any value in putting a low value cap (ceramic? MKT?) across these overheating caps? Would that soak up some of the high frequency ripple?
If so, any suggestions for a value? 0.1uF maybe?
Yes, but the ripple that heats the capacitors up is not that high in frequency, it's the f_sw (and obviously, being a square wave, a few next odd harmonics but the further you go, the less energy there is).
And there is a lot of energy being transferred. A 0.1uF capacitor does nothing; it might provide a tiny EMI improvement in tens of MHz range but that's all.
But you could try some 10-20uF for starters. That could take a non-insignificant part of the ripple off the shoulders of the elcap.
Modern switchers use high enough f_sw (say, 1 - 1.5 MHz) so that 10-20uF really suffices, enabling MLCC-only solution.
It's a good idea to keep the elcaps because they effectively dampen any ringing on the input (for example, caused by power wiring inductance).
-
#11 Reply
Posted by
thm_w
on 10 Aug, 2021 01:03
-
-
#12 Reply
Posted by
daxliniere
on 10 Aug, 2021 19:17
-
I have two of these boards in front of me and both just use a simple linear regulator.
Same regulator as yours EZ1084B 5A 7V max: https://pdf1.alldatasheet.com/datasheet-pdf/view/42415/SEMTECH/EZ1084B.html
They have the section for the buck converter but its not populated. Then a 4 pin molex connector for power, with an unpopulated 6 pin. So its just taking 5V from molex (via the include SATA -> molex adapter), then regulating it to 3.3V.
Maybe you can just convert yours into the 5V input version, or buy another model, they are very cheap. Depending on how much power you are pulling from 3.3V.
Oh man, that is brilliant. Thank you SO MUCH. I looked at quite a few models before I bought the two that I tested, but didn't see that version. I've just ordered a pair of them, I'm sure they'll solve the heat problem. I guess yours don't heat up or you would have mentioned it?
Still, I have no idea why they would use a buck converter for 12->5v, then linear reg for 5->3.3v instead of a) linear reg 12->5v, then linear reg 5->3.3v or b) a 3.3v regulator capable of accepting 12v input.
-
#13 Reply
Posted by
bson
on 10 Aug, 2021 19:51
-
Still, I have no idea why they would use a buck converter for 12->5v, then linear reg for 5->3.3v instead of a) linear reg 12->5v, then linear reg 5->3.3v or b) a 3.3v regulator capable of accepting 12v input.
At 3A a 12 - 5 = 7V drop would require dissipating 21W. In fact, the regulator would produce more heat than the load!
(The obvious solution is a 12V to 3.3V buck regulator. Done.)
-
#14 Reply
Posted by
thm_w
on 10 Aug, 2021 22:55
-
Oh man, that is brilliant. Thank you SO MUCH. I looked at quite a few models before I bought the two that I tested, but didn't see that version. I've just ordered a pair of them, I'm sure they'll solve the heat problem. I guess yours don't heat up or you would have mentioned it?
Still, I have no idea why they would use a buck converter for 12->5v, then linear reg for 5->3.3v instead of a) linear reg 12->5v, then linear reg 5->3.3v or b) a 3.3v regulator capable of accepting 12v input.
Looking at the board again, the assumptions made in this thread were incorrect, at least for my board. What happens is:
- 12V goes into the DC/DC buck, output goes to 3.3V rail
- 5V goes into the linear reg, output goes to 3.3V rail
So the linear reg and the buck converter are in parallel. This is to give you the option of either supplying power with 12V or 5V.
They did not diode or the 12 and 5V together, for whatever reason (diodes would have to be very large)?
Anyway, what this probably means is you can use your existing board and power it from 5V via the 4 pin molex, and make sure no 12V is coming in there, or that the pin is not connected.
But, the linear regulator will still get hot, depending on the 3.3V load. Shark IC is something like 0.3A at 1.1V. Assuming that is fed directly from 3.3V, then ~2.4A * 3.3V = 8W. Which matches your spec.
Now the problem is the linear regulator drop: 5-3.3V * 2.4A = 4W of dissipation.
I'm not sure if the PCB is enough to dissipate 4W. Regulator may need additional heatsinking. I'll test this tonight out of curiosity.
-
#15 Reply
Posted by
thm_w
on 11 Aug, 2021 04:52
-
PCB is definitely not capable of dissipating 4W at reasonable temperature.
So the options I can think of:
- change the caps to ultra low ESR high temp versions (as mentioned)
- supply the 3.3V from an external buck converter, and bypass all the onboard circuitry. eg 50W+ rated or something overkill
- large heatsink on the linear regulator
-
#16 Reply
Posted by
daxliniere
on 11 Aug, 2021 15:38
-
Thanks again THM_W, that's great info to get.
I just measured the caps on the 12V rail and I get ~900uF (4 in parallel), but the ESR is only 0.01R
That seems quite low (good!), so I'm not sure if adding lower ESR capacitance will solve anything in this case.
Out of the 2 boards I have, the one with the 4-pin Molex doesn't connect to the 3.3v reg. I have another model on order without the buck converter, so we'll see how that one fairs. I suspect a heatsink for the 3.3v reg is going to be the best solution. Maybe still worth getting a variable buck converter, setting to 3.3v and running from the PSU's +5v line.
-
#17 Reply
Posted by
daxliniere
on 11 Aug, 2021 16:50
-
Righto! Adding extra caps made a difference to the overheating cap situation, but of course nothing for the overheating 3.3v reg.
I did some more digging and it seems that you can also get these riser boards with SATA power connectors on them. VERY PROMISING!!
All of the ones I've seen have the buck converter unpopulated, but they still have the 3.3v reg. The really silly thing is that the SATA power standard includes a 3.3v line! (though some PSUs don't actually have this as 99% of drives don't actually require it.)
I have tested both of the PSUs that I intend on connecting the DSP card to and they both have 3.3v in the SATA connector. The manual for my BeQuiet! 550W Straight Power 11 PSU indicates a single 3.3v supply rated at up to 24A. (The 4x 12V rails are all listed separately, so it seems logical that the 3.3v found in the SATA power connector is the same 3.3v that runs the motherboard, as opposed to a separate SATA-only rail.) Couldn't find much info for my Kingwin STR-500 PSU, other than that the 3.3v rail is rated at 20A.
My plan is to cut the traces that lead from the 5V input to the 3.3v reg and simply bypass the whole thing. I might wire the caps in. Can't hurt, I guess??
-
#18 Reply
Posted by
daxliniere
on 27 Aug, 2021 20:40
-
Okay, so with a SATA-power version of this riser board (VER 007S), I cut the 5v trace to the input of the 3.3v reg, removed the reg and connected directly to 3.3v from the SATA power connector.
The card was seen by Windows device manager, but the DSP card's software couldn't see it. An old firewire card I had lying around seemed to be detected perfectly also (but had nothing to test it with.)
I reconnected the 5v trace to where it was originally, but with the 3.3v reg disconnected, it ensured no 'crossed wires'. Now the card worked, but under heavy load, it seems to drop out and the software can no longer communicate with the DSP card. Hmm. I inserted a DMM to measure current on the 3.3v line; 1.7A idle, 2.25A under max load.
Testing with another one of these 007S riser boards had the same problem, so I am starting to suspect the riser board, rather than my connect-to-SATA-power experiment. I got 6 of these risers, so I'll test a few more to see if I can find anything out.
After that will be continuity tracing the riser to see what's going on. If you recall, the 009S and 006S risers worked flawlessly except that they got way too hot to be a long-term solution.
It's all very strange, but I'm still bent on finding a solution!
-
#19 Reply
Posted by
daxliniere
on 27 Aug, 2021 20:53
-
I'm now thinking I might go back to one of the VER 006S boards (molex 4-pin) which were working correctly and implementing my own 12->3.3v conversion using a more efficient device. This is what I've found:
https://uk.rs-online.com/web/p/switching-regulators/7813282/Not the cheapest, but it can cope with 3A, which was hard to find.
Any thoughts on this?
-
#20 Reply
Posted by
thm_w
on 27 Aug, 2021 22:35
-
-
#21 Reply
Posted by
daxliniere
on 29 Aug, 2021 13:05
-
Thanks @thm_w
I have some of that kind of thing (LM2596S based), but wouldn't that be just as bad as the on-board buck converter/linear reg situation? I would be happy to use one of that type, since the TRACOPOWER module is over £30, but worried it wouldn't be an improvement.
-
#22 Reply
Posted by
thm_w
on 01 Sep, 2021 20:23
-
Some will be built better than others, I really don't know what caused the original heating it just seems like bad design somewhere.
Here is a photo of the item in the second link running at 12V in 5V out, 2.5A out.
The diode gets warm (70C) and everything else is stone cold.
-
#23 Reply
Posted by
daxliniere
on 26 Sep, 2021 13:04
-
Hey @thm_w,
Thank you for the fantastic suggestion, those little "Mini 560" buck converters are awesome! I went back to the VER006S riser board (4-pin Molex power) as the VER007S (SATA power) boards I got literally did not work. Removed the 3.3v linear regulator, installed the Mini 560 (powered from the 5v line) and changed the CapXon caps to Nichicons of slightly lower capacitance but higher voltage rating. (It's what I had on-hand.) Yesterday it passed over 6 hours of stress-testing with minimal heat generation. I'd call that a big win!
And thank you everyone for your help and suggestions.
All the best!
Dax
-
#24 Reply
Posted by
daxliniere
on 26 Sep, 2021 13:09
-
Next steps are to design a 3D-printed case and to work out what power connector I want to use. Gotta be 3-pin and small enough to fit through a single computer card slot.