RG1 1.8v regulator

I just did some digging in the NCP1117 data sheet (ONsemi version).

RG1 is an NCP1117-1v8 fixed linear LDO regulator.  We have suggested that if RG1's regulation voltage Vreg is greater than IC3's (LAN9512) internal Vreg, then RG1 will supply 1.8V current instead of IC3, reducing IC3's power consumption.

The NCP1117 has a "adjustable output" version which sets Vreg using external resistors instead of internal resistors like the NCP1117-1v8.  In addition, you can adjust Vreg of the "fixed" version with an external resistor plus a stabilization capacitor.  Take a look at Figures 31 and 32.  In Figure 31, a 50 Ohm resistor between the NCP1117's GND pin and circuit GND shifts its output by 300 mV.  Figure 32 uses a variable resistor.  Shifting RG1's Vreg up by 90 mV (+5%) might do a nice job of cooling down IC3 -- probably a 15 Ohm resistor.

Figure 31 also confirms that if two NCP1117's are connected in parallel, the LDO with higher Vreg does switch off the LDO with lower Vreg.

Figure 31 also confirms that if two NCP1117's are connected in parallel, the LDO with higher Vreg does switch off the LDO with lower Vreg.

The engineer in me asks which electrical characteristic on the datasheet defines the voltage differential required to turn the regulator off? 

Hi all, first post from another refugee...

selsinork wrote:

 

Figure 31 also confirms that if two NCP1117's are connected in parallel, the LDO with higher Vreg does switch off the LDO with lower Vreg.

The engineer in me asks which electrical characteristic on the datasheet defines the voltage differential required to turn the regulator off? 

It's simply the way that the average linear regulator works (caveat alert: there's no schematic in my 1117 datasheet, so I'm assuming a standardish circuit). A linear reg (very basically) dumps current into a load while monitoring the voltage on it's output pin with reference to it's ground pin. A feedback loop varies the current from input to output to keep the output pin at setpoint. If you apply a voltage higher than setpoint to the output pin via a low impedance source the regulator has no way of dumping current from output to to ground internally, so the output voltage is pulled high. The internal error correction then takes a dump (technical term!). So then in this case, the regulator with even a marginally higher output voltage supplies all the current to the surrounding circuitry, while the other is shut down. As has already been mentioned there may be parasitic oscillation, thermal effects... allsorts going on too. There's no easy way to make fixed regulators work in parallel either - best to let the right one do all the work! What's the upper voltage limit of the 1V8 line anyway?

Hope this all makes sense.

There's a report today that adding a heatsink cures networking problems:

http://www.raspberrypi.org/phpBB3/viewtopic.php?f=24&t=14478&start=4

by lajos » Thu Aug 16, 2012 3:01 pm

So I did an experiment. I installed heatsinks on both the SOC and memory chip. It's a small heatsink, the SOC still gets pretty hot (I don't have a surface thermometer, but it's not comfy to the touch.) But now the ethernet works without dropping out.

 

Before heatsink, I would get at tops 40-50KB/s transfer rate, with frequent stalling, with heatsink I can move large files at ~2.0MB/s.

 

Although there is a contrary report as well:

http://www.raspberrypi.org/phpBB3/viewtopic.php?f=28&t=12097&start=247

I added a radiator to the chip but there was no difference. Perhaps it was too small.

 

That is sort of a synonym for "unreliable."

-J

Hope this all makes sense.

Possibly, but everything you say is based on assumption with no actual reference to defined behaviour.

There was a very simple reason for asking that question. If the behaviour is not defined you have no basis to rely on it, you simply don't know if it'll work or not. Saying "It's simply the way that the average linear regulator works" is not an acceptable substitute for actual defined behaviour.  Doesn't anyone question why the behaviour isn't defined ?

The only reference appears to be Figure 31 and that suggests you need 300mV to (reliably?) turn the regulator off. So how does the SoC, RAM etc react 2.1v on it's 1.8v input ?

On the other hand, if it's only a couple of mV then you'll have lots of fun with oscillations as the voltage measurably changes by a few mV depending on how you load the cpu/gpu.

Without a definition of the behaviour it's only ever going to be a bad engineering decision to use it, rely on it, or expect it to work.

Also we don't know what the 1.8v regulator inside the lan9512 is or what it's defined behaviour is, it could be a 1117 clone, or could be something vastly different in design, we simply don't know. We do have an answer from SMSC saying it's not designed to be used this way.

selsinork wrote:

 

Hope this all makes sense.

Possibly, but everything you say is based on assumption with no actual reference to defined behaviour.

 

There was a very simple reason for asking that question. If the behaviour is not defined you have no basis to rely on it, you simply don't know if it'll work or not. Saying "It's simply the way that the average linear regulator works" is not an acceptable substitute for actual defined behaviour.  Doesn't anyone question why the behaviour isn't defined ?

 

The only reference appears to be Figure 31 and that suggests you need 300mV to (reliably?) turn the regulator off. So how does the SoC, RAM etc react 2.1v on it's 1.8v input ?

On the other hand, if it's only a couple of mV then you'll have lots of fun with oscillations as the voltage measurably changes by a few mV depending on how you load the cpu/gpu.

 

Without a definition of the behaviour it's only ever going to be a bad engineering decision to use it, rely on it, or expect it to work.

 

Also we don't know what the 1.8v regulator inside the lan9512 is or what it's defined behaviour is, it could be a 1117 clone, or could be something vastly different in design, we simply don't know. We do have an answer from SMSC saying it's not designed to be used this way.

Very valid points, except perhaps the "fig 31...suggests you need 300mV to (reliably/) turn the regulator off" part. "Yes... and no" is probably the best answer I can give to that.

Who's to say that the nice folks at On Semi ever actually built and tested the circuit, rather than just lifting it from a cookbook and specifying a nice big differential voltage to allow for the vagaries of production tolerances, changes due to ambients and a lump more for safety? And on an unrelated note,  why would a "proper" engineer specify 50R instead of a more sensible (from an inventory POV) 47R? I see this all the time on application notes...

The fact that On specify applications with a resistor in the ground lead demonstrate that (like any other vanilla linear regulator) this one does not rely on a mechanism for dumping current to ground, therefore any overvoltage applied to the output above setpoint will be resisted by the feedback circuitry in the only way it knows how, i.e. by increasing the resistance of the series pass element. How a pair of regulators fight over current supply duties in this real world case will be more complicated (and will require real world analysis), but I'd bet a night's beer money given the output voltage regulation spec that it would take rather less than 300mV output overvoltage to switch one off. The actual setpoint may vary a tad due to the reasons above, but once it's exceeded by even a small amount, the regulator stops working, because there will be a high gain under the feedback loop. It all depends very much on the characteristics of the other reg too, which we're ignoring...

The bottom line is of course that it's not supposed to work like this and the usual "we'll work around it in software" hacks won't help. I really can't believe that they're still churning these things out at full pelt...

Very valid points, except perhaps the "fig 31...suggests you need 300mV to (reliably/) turn the regulator off" part. "Yes... and no" is probably the best answer I can give to that.

Well as it's not defined elsewhere, yes and no is about the best answer I'd expect   Taking a random sampling of datasheets from other manufacturers of equivalent parts, some include that circuit, some don't. I didn't find another one that mentioned the 300mV or defined this behaviour at all.  What is clear is that there must be an original manufacturer and they're all mostly cut&pasting the datasheet from one source. I couldn't easily work out who the original designer was so it's hard to pick any one of the datasheets as being authoritative.

And on an unrelated note,  why would a "proper" engineer specify 50R instead of a more sensible (from an inventory POV) 47R? I see this all the time on application notes...

I'd probably have picked 51R, but as 50OHM, 0.1W, 1% 50OHM, 0.1W, 1% and all sorts of strange values like 49.9R, 50.5R, 51.1R are available at the same cost as the 'standard' values. I agree though, it's a pain when somthing specifies some odd fractional value that you have to order up specially.

but I'd bet a night's beer money given the output voltage regulation spec that it would take rather less than 300mV output overvoltage to switch one off.

So would I.. The point was simply that we don't know and as there's no reference we're simply not in a position to judge whether this is suitable behaviour or not. The engineer side of me does things like "specifying a nice big differential voltage to allow for the vagaries of production tolerances, changes due to ambients and a lump more for safety?" as I suspect most of us do, unfortunately that's exactly what brings us to the 'undefined behaviour' vs real world application question.

The designer of this battery changeover ciruit (ON Semi's fig 31) was intending to set a difference so that one regulator would do all the work if the ac power was present. The 300mV (which is obtained by assuming the nominal 6mA quiescent current flowing through the 50R resistor) gives a margin over the +/-100mV tolerance of the regulators' output voltages. It's not  a good design because the minimum quiescent current is not specified so the intended 300mV might easily be only 200mV and then a worst case pair of regulators would be sharing the load.

What we do know is that the 1117 regulator can be used in pairs but since the SMSC chip makers explicitly say that their part should not the RPi circuit is still uncharted territory (and obviously just a mistake).

I can't believe that they don't just put it right.

Michael Kellett

which is obtained by assuming

See, that's the bit I don't like

I can't believe that they don't just put it right.

Hopefully they will, but for now all we have is another assumption that theres no quick and easy way to do that.  If there's a pcb respin needed then it makes sense to take some time and fix as many problems as can be identified in one go - alignment of the RJ45 would be quite high on my personal list. 

There's likely to be some number of bare pcb's in the pipeline, so even an immediate fix could take some time to filter through to customers. We have no details of any of that, nor would I expect to ever see it, so we're left with more of those pesky assumptions

Hi selsinork and thanks for the reply. You're absolutely right  - no amount of assuming on our part is going to move any copper! Putting on my optimistic to the point of naïveté hat for a moment then if a board revision does come it will take some time to proto and test (properly!) - way longer than the factories' stock of the current boards (which will be pretty skinny). Raspberry Towers may be wary of moving that connector (or doing anything else that isn't absolutely imperative) given the amount of fubar already. Who knows with that lot? They certainly make assuming stuff difficult. I do detect a little more humility on the forums over there lately though. Sometimes. With certain staff. When no-one's looking. Most of the fanbois are just as obnoxious as ever though.

I wonder whether the rollout (if it does happen) will be accompanied by fanfares, or whether it will be a subdued affair...

Hi selsinork and thanks for the reply. You're absolutely right  - no amount of assuming on our part is going to move any copper! Putting on my optimistic to the point of naïveté hat for a moment then if a board revision does come it will take some time to proto and test (properly!) - way longer than the factories' stock of the current boards (which will be pretty skinny). Raspberry Towers may be wary of moving that connector (or doing anything else that isn't absolutely imperative) given the amount of fubar already. Who knows with that lot? They certainly make assuming stuff difficult. I do detect a little more humility on the forums over there lately though. Sometimes. With certain staff. When no-one's looking. Most of the fanbois are just as obnoxious as ever though.

I wonder whether the rollout (if it does happen) will be accompanied by fanfares, or whether it will be a subdued affair...