is there an opportunity to do undervolting Mediatek, Marvell SoC, etc.? is there ready-made software for this?
I'd say most of the converters are fixed voltage output.
You're welcome to mess with the feedback circuits? =P
As an example, having a look the best place to start might be have a look at
https://one.openwrt.org/hardware/
? I can't see much in clock and power control for any sort of digital interface between CPU and the converters.
What is undervoltage supposed to accomplish, more than brownout and master reset on a SoC?
1 Like
For street credit? Lower temps? So they can mess with the clock control later? Fun =P
I don't think this is a good idea either.
But I guess one could try to tune everything to the minimum voltage recommended under the MT7981B datasheet or further just because =P They have +-5% on the rails according to https://one.openwrt.org/hardware/MT7981B_Wi-Fi6_Platform_Datasheet_Open_V1.0.pdf
Or set everything to the maximum and add cooling and then try to mess with the clocks =P
that is, I will still be able to control CPU voltage on chipset if I replace resistor with a variable one that I can change myself resistance Mediatek SoC. Only according to SoC documentation do I need to find out if it has 3.3v and 5v pinout, right?
well, I can try to measure it with DCV multimeter nearby SoC without documentation, most likely these resistors will be near to power phases (power inductors) as on laptop motherboards.
Yeah. "just" need to reverse engineer the converters. and "find" all the rails =P Also note that some of them may be LDO's and not switching converters. And some of them will be fixed output with internal references. So that would require more significant modification.
Better to just go for things with documentation. But yeah most converters you'll be able to find documentation for. Less NDA's than with CPU's / networking gear. As stated above I don't see the point unless one just wants to learn about electrical engineering =P
Task is to increase service life of the platform, because most often SoC itself wears out, possibly from overheating.
Hahaha. I would be more worried about capacitors.
Add more cooling.
If your platform has LDO's for the main big voltage rails look at putting a switchmode converter on it instead to get less heat in the box.
Going for small gains with voltage mods not worth it.
Yeah, electrolytic capacitors are the only components that could 'wear out' over time. If anything, keep an eye on those.
1 Like
Yeah. The number of things I see with 85 degree or "1000 hour" caps "because cost" on them near hot components or with no airflow/sealed applications is always funny =P
1 Like
After reviewing datasheet specification my SoC, I was convinced that it is possible to achieve a guaranteed voltage drop to -5%, everything below already needs to be experimented with and depends on purity of the silicon substrate during lithography.
You're not taking into account transient behaviour.
For LDO's: I haven't quickly found a good document for switching converters yet:
Something about switching converters:
Video:
Routers (and x86/SBCs) are actually quite robust and can run for years -- 10 years or more is not uncommon, and usually to the point where the hardware is simply not powerful enough to run a modern version of OpenWrt.
With that in mind... Leave the device alone and add some cooling. A small desk fan blowing over the device can be surprisingly effective.
You're more likely to decrease the service life to ~0 hrs when you modify the hardware and break it.
7 Likes
Like I said, burnout is coming.
But the voltage you have marked is core voltage.
Does your SoC in mind have internal core voltage regulation which is the standard and only Vcc as power supply voltage since I/O voltage is referenced to Vcc.
But this “datasheet” doesn’t look like a complete datasheet for design use. This is only a sales leaflet.
3 Likes
But you can’t compare a 100+W PC/Laptop APCI power control with a lifespan of 2-3years with a router SoC and its 10-20W power supply made for 24/7 work load for maybe 10years.
2 Likes
That looks like it comes from the 54/52 page "datasheet" one finds by google =P It doesn't specify transient response/ noise / regulation etc for any of the rails. It also doesn't give a reference implementation on the power supplies either =P
As I said before. Can't just look at DC. You need to take into account the transient characteristics. There's a reason why there's all sorts of brownout protections in those big modern big CPU's and they have fancy converters that talk to the CPU. We're talking dumb fixed voltage output converters and LDO's on this gear.
But it does give the pinouts, power on sequence and DC electrical characteristics for all the rails.
3.3 +-10%
Vcc15 -13%+8%
RGMII 2.5 rail -10%+8%
RGMII 3.3 +-10%
DDR1 +- -4% + 8%
DDR2 +-5%
3.3V rail is 218-436mA
1.5v rail is 147-173mA
1.27v rail is 380-540mA
DDR2 current is 95-253mA.
So for example design something that can handle a 218mA to 436mA spike and/or load release whilst not dropping below 3.0 volts or going over 3.6v =P
For example for a reference implementation of a 4 phase buck converter operating at 1.2MHz you might be able to get to 1.2375v if you expect +-0.5% regulation and a no load voltage of 1.2175v. How much effort for 32mV? (Not to mention switching converter so you might have 10-20mV of ripple there =P) Undershoot might be 150mV if you don't have the current sense network tuned properly =P I haven't even talked about thermal compensation yet =P
Another converter: switched capacitor gets you +-3% and load transient of 25mV with a converter that can do 10A.
Transients are all about time vs how fast the PSU can compensate for loads and the capacitance and resistance for the PCB wires and the buck capacitors that compensate for this transient time factor.
But the problem will almost always be current rushes that bring the voltage down and cause burnout. If the current goes down the overvoltage transient from psu will get stuck in the capacitors trying to charge them.
So to have normal operations on Vcc min will become brownout conditions sooner or later.
I assume this will be in a home with stable temp. Not operating between -40°C to 125°C. Because then this gets even more complicated. Especially below 0°C.
1 Like
Mm. Yes so load release and load ramping up. We don't have those specifications. Like an application note on bypassing etc =P.
mm. You need to make sure the converter can react in time. Depends on current sensing and the reaction time of the converter.
This is also about sizing the bypass capacitors as well as how you design the filter network for the feedback. Had a look at a cheap 1A LDO datasheet and it you're still talking -100mV from a 1mA to 1A load transient and then starting to recover. Then load release gives you a 100mV jump the other way and then it starts to recover. (Cout being 1uF)
You also have inductance and capacitance between the output of the converter and the load yes.
Mm. Had a quick look at some 1.2v LDO's and they seem to be temperature compensated within 0.1% over 0-75 degrees C. But still we're talking changes in voltage.
Even a "stable" household temp might vary 10-20 degrees? How stable do you need it to get? If we're talking right down to the line as OP stated where they want 1.20 volts exactly at DC =P
The thing isn’t to buy components that have the right numbers. The thing is to make them actually work together as a complete system.
1 Like