Let's move the moulds!

[levi]

It's logged on a 15 minute granularity at present, but yeah, if someone pre-orders an EU cellular 4GB unit at the same 15-minute span that ED is processing some cancellations of the same unit then it's impossible to differentiate that from a smaller number of cancellations. I'm not sure how often that happens - people would need to hit the same 15-minute slot and also preorder the same variant of unit as ED is cancelling, either non-cellular, US-cellular or Euro, and independently either 2GB or 4GB. I guess it's most likely to happen with preorders of 4GB EU units, as most pre-orders that come in are of that type still, and cancellations seem to come in to that category in larger numbers than the other 4GB categories (we haven't had a 2GB part cancelled for a while as I remember).

It also generally confuses the hell out of me when someone pre-orders in the same 15-minute moment that ED processes some cancellations when I'm trying to write up the weekly stats report, but that's my problem and not anyone else's.

 

[ible]

don't pick on Zwerg01 too much, as he could be at least 90% right. if ED can't deliver on 4GB, there may be more cancellations than would be expected with an optimistic fanbase.

 

[Silent-Hunter]

Well I should hope not. If this kills the Pyra I'm going to be really mad at the people who just couldn't accept that 2GB is enough.

 

[dracmas]

As hard as they are working on getting 2gig/4gig done I'd say it's more a matter of time till they get all the kinks worked out. Might be a big deal if you want it asap, but I think it's still too soon to be worried about not getting a pyra

 

[Dark Pulse]

Well I should hope not. If this kills the Pyra I'm going to be really mad at the people who just couldn't accept that 2GB is enough.

I doubt this would kill the Pyra, since the issue isn't "it's not working period," it's that the design that was good enough for 2 GB isn't robust enough for 4 GB.

They'll tweak it, and it'll work out just fine. Just might take a bit longer.

 

[levi]

As hard as they are working on getting 2gig/4gig done I'd say it's more a matter of time till they get all the kinks worked out. Might be a big deal if you want it asap, but I think it's still too soon to be worried about not getting a pyra

I quite agree. The last messages on the kernel mailing list from Nikolaus sounded like he was quite confident of an improvement in voltage regulation for the new boards. It just slightly annoying that there have been no posts on that front to the mailing list for two weeks now, but I guess these boards can't get made and populated and quality checked overnight.

 

[Hellfleck]

There is one _very_ important outcome of getting the 4GB boards to work reliably : it will improve system stability for 2GB units as well

Electronic components have tolerances which means their values can swing both ways to a certain amount and electronic components age which alters their values as well and there is heat and other external influences... now imagine the 2GB boards running reliably but just barely so with a thin margin for error, given time these units may become very unpredictably unstable in the future and that would be an absolute mess - no one can tell for sure right now but having a higher margin for those errors is a big plus in my book even on the 2GB units

 

[Silent-Hunter]

That is a very good point. Even if I choose not to ever have a 4GB board, I might want a new 2GB one.

 

[Mr_Loon]

Well I should hope not. If this kills the Pyra I'm going to be really mad at the people who just couldn't accept that 2GB is enough.

I don't think there is any chance the 4GB problem will kill the Pyra, might leave a few flesh wounds mind you.

Sorry to sound like a stuck record here but what I do believe would be genuinely useful at this stage would be some side by side comparison of stable 4GB and 2GB units showing the impact of swap performance & battery life. This would really help people to choose between the options available.

From my experience with a cheap windows tablet that uses Nand for swap is that swap performance will be fine with the 2GB model but who knows until tested?

On a side note can anyone explain why the stability issue only impacts the 4GB boards, is it simply that tolerances are greater on the ram modules used on the 2GB ram boards? Or is the 'signal' somehow noisier when powering 4GB of ram?

 

[Letalis Sonus]

On a side note can anyone explain why the stability issue only impacts the 4GB boards, is it simply that tolerances are greater on the ram modules used on the 2GB ram boards? Or is the 'signal' somehow noisier when powering 4GB of ram?

They need more power, and a higher amperage means a higher voltage drop on a resistor such as the supply line itself. And a lower supply voltage means being closer to the lower voltage limits...

 

[Dark Pulse]

There is one _very_ important outcome of getting the 4GB boards to work reliably : it will improve system stability for 2GB units as well

Electronic components have tolerances which means their values can swing both ways to a certain amount and electronic components age which alters their values as well and there is heat and other external influences... now imagine the 2GB boards running reliably but just barely so with a thin margin for error, given time these units may become very unpredictably unstable in the future and that would be an absolute mess - no one can tell for sure right now but having a higher margin for those errors is a big plus in my book even on the 2GB units

I made much the same point recently myself, and as I said at the time, this could be a blessing in disguise.

Either way, it's not the end of the world in Pyra land. Yes, the feature creep (as some call it) does mean that the design didn't hold up when 4 GB of RAM became an option. But seeing as how the difference between "working perfectly fine" and" not even making it to boot" were (essentially) a doubled consumption and power draw from the RAM, making that aspect more robust means that it will become much more possible to, say, overclock your Pyras - whereas before, 1.5 GHz may have been, quite literally, its limit.

 

[Zwerg01]

They need more power, and a higher amperage means a higher voltage drop on a resistor such as the supply line itself. And a lower supply voltage means being closer to the lower voltage limits...

It is already proofen that this theory is wrong. hns soldered in additional wires to reduce resistance with no effect.

 

[levi]

There is one _very_ important outcome of getting the 4GB boards to work reliably : it will improve system stability for 2GB units as well

I thought that too for a while, but then I considered that to test these boards they're running some sort of memtest on them, which would I thought have stressed the memory power usage far above what it would doing almost anything else. What I've snipped about parts ageing may be valid still, but if they're all passing at present it would suggest that most of them are well above the line when running memtest at least. Just one row going bad marks RAM as being basically unusable going forward of course, but we're likely looking at intermittent failures here I think when RAM stops working normally. I've lived with intermittent failures for many years on machines I've owned in the past.

 

[Letalis Sonus]

It is already proofen that this theory is wrong. hns soldered in additional wires to reduce resistance with no effect.

High frequency electronics are affected by more than plain static resistance in the same way the coil of an electric motor always has a much larger effective resistance while the motor's spinning. A bypass wire may as well just act as an additional antenna sending noise to nearby lines or turn into a situational resistor by receiving them. Even worse, we're talking about frequencies where the length difference of separate lines that depend on each other matters, a separate wire whose length does not fit to the line it is bypassing may as well cause noise to peak even more due to temporal interference.

There's a lot more physical background behind this to just call it entirely disproved by such a simple test, and I'm sure hns knows this. There's a reason someone already helped out with expensive-as-fuck simulation software before they even ran into this issue.

 

[T.T.]

In order for the bypass wire to make any useful difference, it would probably have to be 2-3 mm thick just so the inductance can be low enough to not mess with the frequencies we're dealing with. (and even that may not work due to skin effect. Litz wire may work though)

 

[Zwerg01]

we're talking about frequencies where the length difference of separate lines that depend on each other matters

No, we don't do so. The wavelength is approx. 56 cm wheres the trace lenth is only a few centimeters and i am pretty sure whoever made the layout followed the design guidlines. Everything else would be damn stupid. A copperwire would decrease resistance cauased by either skin effect or trace dimensions dramatically even if it is a few centimeters longer. Sorry, but you mix up everything and out comes only poo-poo.

 

[Letalis Sonus]

The wavelength

...doesn't matter much on the noise that may already be present from other components, don't ignore the rest of the sentence. If there's enough time difference on two lines to make one noise peak overlay the other, you already have a problem. These peaks do not have to be directly related to the component they are going to, the RAM is not the only high frequency component and it is not being driven synchronously to those others.

I'm pretty sure you calculated the wavelength of the clock signal, not the actual IO bus - cut that in half. Within copper the wavelength goes down a by a good chunk as well and it takes a lot less than a full period to confuse the circuity, you reach temporal complications earlier than that if you only look at the direct circuity. Also, only a fracture of the wavelength is necessary to get an antenna for a certain frequency, there are actual passive antennas being used with only 1/10 in length - being closer to the wavelength just allows to increase the efficiency - a copper line is already affected with a much shorter distance than that. You are not using over 3m long antennas for your FM radio, are you?

 

[daveshah]

The reason the length difference matters so much in the case of trace length matching is not so much to do with the frequency or wavelength but because of the requirement specified in the interface for the data to arrive within +/- x picoseconds (IIRC 25ps for the data bus) of the strobe or clock signal.

In actual fact even if the interface was operating at 1Hz if the interface required data to arrive within a 25ps window of the strobe then it would still require just as much length matching as a 1GHz interface. Of course systems running at 1Hz would be designed with a lot more slack, but it's interesting food for thought.

This is probably not relevant for the issue at hand, which is the power supply not data routing. In this case the most important thing is keeping the impedance low. In many cases this is more about decoupling capacitance and ground plane integrity than trace thickness, so I can easily see a wire not helping. A power plane - which could be seen as a thick trace - is usually used below and around the memory, particularly on a high layer count board. This ensures the lowest possible impedance at high frequencies near to the memory device, as not only is it a significant area of copper but it also benefits from distributed capacitance with an adjacent ground plane. There will also be a number of small decoupling capacitors near the device, again to provide a low high frequency impedance at every power ball of the device.

If there is a problem close to the memory device - such as an inadequate power plane or decoupling capacitors, then a wire wouldn't help at all. Remember that every memory device has a number of power supply input balls, and there are 4 devices. That adds up to a lot of points each of which must have sufficiently low impedance, and at high frequencies that can pretty much only be done with good planes and decoupling.
[doublepost=1509790374,1509789211][/doublepost]To add some numbers: the typical supply impedance target for DDR3 might be to be well below 1 ohm at 500MHz (I've seem target values down about to 0.1 ohm).

A 3cm wire has an inductance of about 24nH and won't have significant capacitance to ground, meaning an impedance of about 75 ohm at 500MHz.

While the wire test is useful to rule out issues such as an inadequate voltage regulator, it does little to rule out higher frequency issues.

 

[Swordfish II]

They need more power, and a higher amperage means a higher voltage drop on a resistor such as the supply line itself. And a lower supply voltage means being closer to the lower voltage limits...

This. There is no improvement for the 2GB by fixing the 4GB.

It is like building a bridge that can support a million tons when it can already support 200 tons, and will actually only see 100 tons.

 

[elwing]

This. There is no improvement for the 2GB by fixing the 4GB.

It is like building a bridge that can support a million tons when it can already support 200 tons, and will actually only see 100 tons.

Not fully true, lower voltage drop can mean better overclockability for the 2GB... Also the 2GB is deemed stable as is, but we won't be sure they won't get some that won't be stable until mass production...