RG1 1.8v regulator

As shown in Troy Mackay's post on Jul 28 it seems to me that we could mod our PI's to work more as the chips were designed.

I think that this could make the PI more stable, from the looks of it.

I am not a designer Just an old tech, but I think we need to find a way to FIX what we now know is an error in the board.

So could someone do some practical testing to see if there is something that can be done to easly fix the current board, rather than wait for RPI foundation to fix it by waiting for a board redesign?

i will love to try out the hack Troy mackay has done also and then test again with that fix on the board, but i have looked into this and i most say it is very well done by Troy as i think it is to small for me todo and i done have an microscope as need for this.

so yes if some one can findout where to make an cut to split the LAN9512 1v8 from the lod 1v8 then i will try this also.

Tooms

Everyone doing their best is assumed.  It has no bearing on whether a product works or not, except possibly if the testers are ineffective and hence undermine everyone else on a project.  If anyone else in an engineering team fails, it should be caught at test to within certain limits (not 100%).

But more importantly and relevantly, that issue is not part of the engineering process.  That process at the core of our profession concerns the objects being engineered, not the people.

Here I've assumed that at least since Intel's fdiv bug, hardware

guys have had the advantage over software guys because they

use mostly formal verification techniques where we rely mostly

on ad hoc testing.

Don't get me wrong, formal verification has it's place and can help a lot. It also leads to two sometimes difficult problems

1. a simple error early in the design or even in a symbol/footprint supplied by a third party gets propagated through the design resulting in something that may not function at all, or as in this case works but is suboptimal. As the verification passes everyone has a high sense of confidence it'll work.
2. Over time it leads to a reliance of the verification technology and a de-skilling of the job - beancounters strike again.

In short there's no substitute for knowledge and experience. Experience of making the stupid mistakes and fixing them usually helps a lot.

If you're a software person, you're probably familiar with unit testing.  IME it suffers from variants of the same problems.

1. The spec is wrong, unit test to the letter of the spec and the result is wrong, but the test passes.
2. Often the person who writes the code also writes the unit test for it, if they have some misunderstanding of the requirement, the test will pass but the result is still wrong.

The same observation that there's no substitute for knowledge and experience holds for software just as much as for hardware.

A dodgy datasheet, or software spec, just means it's even harder to spot the error.

Hardware people do ad-hoc testing too, the tools are different and we literally get our fingers burnt - with a soldering iron

  As technology gets more complex, ambiguity has to be reduced

so checking can be increased, just to keep reliability from getting worse.

Agree, but as the complexity grows you're increasingly faced with a datasheet being 1000 pages instead of 50. It's trying to reduce ambiguity, but it's bringing up the needle in a haystack problem.

Then there's always the problem where increasing complexity gives you 17 different ways to do something, none of them obviously better or worse.

Since we're talking about testing now, and looking for ways to improve future versions of Pi, let's focus on that a bit since it's clear that testing needs to be revamped to a significant degree.  This observation stems from the fact that testing somehow managed to miss the dramatically widespread incompatibility with common class-compliant USB devices that users have encountered and reported, as well as not catching any USB event or data loss, nor the high operating temperature of some LAN9512 devices.

It seems that USB testing fell short in several respects:

• Sample size.  Even if devices had been chosen at random, with a sufficient sample size it should have become apparent that there is a large incompatibility problem and that data loss is common.
• Breath of devices tested.  Within the range of devices that fall within Pi's intended application area, a representative selection of device types and most common brands would have raised confidence that Pi works correctly with normal class-compliant devices that use only the generic drivers for their class.  It is very clear that Pi does not.
• Depth of functional testing.  This refers to performance, overhead and error rate characterisation.  USB supports various different types of data transfer, but the most common and important types are lossless, ie. these USB transfers are meant to be 100% reliable.  Users have found that USB data loss from mouse and keyboard are common, but this is hard for users to quantify.  Given the importance of these HID devices, the error rate should have been characterised at testing time.
• Physical tests.  Although only laterally related to USB testing, the wide variation in LAN9512 operating temperatures was not caught during Pi testing.  This suggests that either the test sample size was not large enough to reveal the problem or that it wasn't conceived that temperature measurement should be added to the set of tests that were performed.

Everything is a trade-off, and perfection can never be reached through testing.  However, basic correct operation is non-negotiable, so enough testing needs to be performed to guarantee that each of these categories are covered to a satisfactory degree.  That is the only way to have measureable confidence of engineering success upon product release.

Assuming that the premise is agreed, we then have to come up with a way to achieve this within cost and time constraints.  It has to be agreed first though.

Morgaine.

selsinork wrote:

 

If you're a software person, you're probably familiar with unit testing.  IME it suffers from variants of the same problems.

1. The spec is wrong, unit test to the letter of the spec and the result is wrong, but the test passes.
2. Often the person who writes the code also writes the unit test for it, if they have some misunderstanding of the requirement, the test will pass but the result is still wrong.

Expanding on (2), IMO you should always have different people designing and testing.  A good designer wants to make something work, a good tester wants to break it.  Getting both mind-sets in one person is a rarity, and inevitably leads to a conflict of interest.  Also, humans have difficulty perceiving what's really there versus what they want to see -- you should have someone else proofread your writing, and review your designs.

selsinork wrote:

 

Agree, but as the complexity grows you're increasingly faced with a datasheet being 1000 pages instead of 50. It's trying to reduce ambiguity, but it's bringing up the needle in a haystack problem.

There are good and bad tech reference manuals out there.  It all depends on whether the vendor sees a manual as (1) a way to get more design wins and require less tech support, or (2) an painful burden to be dispatched as cheaply as possible.  A 1000-page TRM isn't a problem if it's well-organized -- I can just print the chapters I need for my particular application.  Many SoCs have all sorts of functions that I don't need, so I can just ignore them provided that the SoC is well designed and there aren't unexpected interactions.  OTOH, some 3000-page TRMs are that way because the writer copied and pasted identical functions and then made small changes to each copy instead of thinking out clearly what would be most helpful to the designer and making one copy with a table of individual differences.

When a manufacturer is dodgy about documentation, it makes me wonder how well their products are designed, since a well-designed product generally starts with a well-written specification, which then provides the basis for a well-written TRM which doesn't cost much to complete.

John Beetem wrote:

 

When a manufacturer is dodgy about documentation, it makes me wonder how well their products are designed, since a well-designed product generally starts with a well-written specification, which then provides the basis for a well-written TRM which doesn't cost much to complete.

A TRM is written as much for the manufacturer's own staff as for the external audience.  If a company can't be bothered to describe to its own staff how its product works, it doesn't inspire any confidence that its staff actually knows how it works, so you can expect support and the company's support products to be poor or non-existent.

Orthogonal to the above, the days of monolithic TRMs or other large specification documents should have disappeared long ago.  Scalability refers not only to systems but to documentation as well.  Any non-trivial document needs to be hierarchical, distributed, and revisioned at each level of hierarchy, or it has no hope of being maintained to track the thing it describes.  Unfortunately today's prefered format for technical specs favours monolithic documentation.

Even if the delivery is "monolithic", internally inside the company the separation should take place.

With "monolithic delivery" I mean: "A single datasheet that describes the whole chip".

With separation I mean that the different chapters are maintained and written separately.

Care must be applied to ensure that when a module is functionally upgraded, the datasheets for the previous versions of that module don't get more complicated.

Having now read the last few pages of this thread....

The LAN chip is not designed to deliver power on the 1.8V. To stick to the specs the two signals need to be decoupled. i.e. a PCB redesign.

In practise the draw on the 1.8V is about 100mA. This means 3.3V-1.8V * 100mA = 150mW of power moves from the LAN chip to the regulator.

In practise, on the total power-draw of almost a Watt, the lan chip will be able to handle the 1/6th of a watt extra power. Maybe not over the full temperature range, so stuffing the 'pi into a mostly-closed enclosure may lead to unneccessary problems.

No recall of existing 'pi boards needs to be done. As said: The lan chip can handle it in practise, it is a question of conforming to the specifications, and being more robust.

If you are comfortable with using chips outside the official specifications, you can save a bit of cost if you leave off RG1. As it seems to work with RG1 removed, or with RG1 there is little current flowing from it. The lan chip seems to be able to cope.

Some people are expressing views that paralleled regulators will have one delivering ALL the current, and the other one none. This is not always the case, This would be the case if they are both ideal voltage sources (with an ideal diode in series). However in practise, both regulators would have a measureable output impedance. So if regulator 1 delivers 1.85V with an impedance of 1 ohm, and the other delivers 1.80V with an impedance of 0.1ohm, the first 50mA will be mostly delivered by the first regulator, but increasing the current draw to 100mA, the second regulator will provide about 50mA of that.

Roger Wolff wrote:

 

Some people are expressing views that paralleled regulators will have one delivering ALL the current, and the other one none. This is not always the case, This would be the case if they are both ideal voltage sources (with an ideal diode in series). However in practise, both regulators would have a measureable output impedance. So if regulator 1 delivers 1.85V with an impedance of 1 ohm, and the other delivers 1.80V with an impedance of 0.1ohm, the first 50mA will be mostly delivered by the first regulator, but increasing the current draw to 100mA, the second regulator will provide about 50mA of that.

The thing is, that isn't how regulators work, they aren't a voltage source with a resistive output.

Regulators, including the two linear regulators we're dealing with here pretty much always consist of a voltage reference with a feedback circuit controlling a variable resistance between the input power pin and the output power pin provided by either a FET or a bipolar transistor. If the voltage on the output pin is above the reference voltage by more than a tiny amount then the feedback circuit will shut off the output transistor and nearly all the power will be provided by the device with the higher reference voltage. Where it gets funny is when the two reference voltages are very close, both regulators provide part of the power and you can get into oscillations and all sorts of fun and games depending on the impedances around the circuit.

A TRM is written as much for the manufacturer's own staff as for the external audience.  If a company can't be bothered to describe to its own staff how its product works, it doesn't inspire any confidence that its staff actually knows how it works, so you can expect support and the company's support products to be poor or non-existent.

You've not worked at any of the places I have then

Reality tends to be that there'll be a warts-and-all document that only the design team in question will have access to. Someone will get tasked to produce a tidied up version which will then get sent to a dedicated team of technical writers who have no real knowledge of the device and somewhere in that loop it'll go to the lawyers before any sort of document can be made available internally, never mind publicly.

You'll also probably have four levels of document, publicly available, available after signing NDA, internally available and design team warts-and-all. So how much substance remains in the public doc ?

It's also highly likely they'll have some sort of records retention policy to make sure anything that could potentially be requested for a court case gets deleted ASAP.

IMHO, the lawyers probably have more input to the public doc than the design team does!

It might be nice to live in an ideal world where everything was done properly, the world we actually live in is more about protecting IP and maximising shareholder profit. In the eyes of the business people that means something completely different to what the engineers want.

To give you an idea, I've worked places where even employees use google to find the docs. Says a lot about the internal organisation..

It gets better, you'll probably find that nobody at SMSC/Microchip will know what the actual capabilities of the in-built regulator are. Their chip designers won't have designed a regulator from the bottom up, they will have included a regulator from a design library, it will have been specified to have a little bit of headroom over what the LAN9512 requires for its own operation. Beyond that we're into "experimental" territory.

It gets better, you'll probably find that nobody at SMSC/Microchip will know what the actual capabilities of the in-built regulator are. Their chip designers won't have designed a regulator from the bottom up, they will have included a regulator from a design library, it will have been specified to have a little bit of headroom over what the LAN9512 requires for its own operation. Beyond that we're into "experimental" territory.