# Simulating DDR3 after routing?

#3

The Olimexino64 uses DDR3 and you can find it as a KiCad project on github and on the “made with KiCad” page.
You might want to look it up for the design choices Olimex made and compare it to your own design.

#4

Thanks guys for the info,
I use a chip from samsung, will it be the same or nearly the same with micron ?
Or you guys have info / application note for samsung chip ?

thanks

#5

It’s 4Gbit Samsung chip
K4B4G1646D

Anyone ?
thanks

#6

This looks like basic comparison of datasheets with KiCad’s footprints etc.
Not a question anybody who is not intimately familiar with both those chips could easily answer.
And there are probably tens of thousands or memory chips made and they become obsoletel weekly or so.

My best guess is you should do you own homework here.

#7

Hehehehe, ok, it’s my homework…thanks

#8
1. Calculate the impedance of the tracks you are using. As long as you didn’t use (too many) via, a simple calculation is enough. The same tool will give you velocity, and what you need for input are board stack up and track width. You can tune that up and down to manually simulate manufacturing tolerances, if you want.

1a. You can use ATLC http://atlc.sourceforge.net if you really think you want a field solver. It’s not hard to use, and it’s free. It will give you impedance and velocity, which is really all you care about in practical terms.

1. Look at the IBIS model of the drivers for the two ends of the interconnect. You can calculate a driver impedance from that. Look up IBIS model to see what data there is in there, but a primer of what it tries to model is http://www.ti.com/lit/an/snla046/snla046.pdf Basically, the data is piece wise linear current vs output voltage when driving.

2. Draw out a typical interconnect. Source, driving impedance in series, transmission line of time delay (length, velocity), termination, parasitic capacitance of the input.

3. Put that into a SPICE deck. It will be less than 10 lines of SPICE.

4. Run a transient simulation.

This first order, lumped element simulation will show you what things are likely to look like, and let you find gross errors. You’ll be surprised how realistic the results look, especially if you get the parasitic capacitance at the input and transmission line (track) impedance close to right. You can model vias as capacitors if you want, just break the transmission line into segments (of the correct length) and put a capacitor to gnd. You might need a 1meg resistance to gnd also, or SPICE will complain about there being no DC path at those nodes… just anything large enough not to affect operation.

I’ve done this with SDRAM and LP-DDR designs, both in co-planer (that is, 2 layers with tracks separated by ground fill) and with multilayer designs. It works fine, the waveforms later validated by 1GHz bandwidth (not sample rate) scopes. Wave shape, undershoot/overshoot, damping, delay all a good match with the simulation.

#9

Thank you very very much for a comprehensive explaination…

#10

What chip is it ?
Can I see the model of it ? thanks

#11

I’ve found :
http://www.y-ec.com/pic/20120726124403318.pdf

Where can I get this file :
i_k4b4g1646b_hyxx1p0.ibs

? thanks

#12

Wow - he knows Mr Google!! (Be surprised by the number of supposedly tech people who don’t know…)

You could call the Samsung sales office near you. They can connect you with a support tech.

#13

Of course I know Mr.Google, but it’s not helping me now…hahahaha
Good idea for calling Samsung, thanks

#14

Is this the same board you were working on almost a year ago? I would just get a few boards made and assemble and test one. Using some “real” equipment on a “real” board would provide far more useful results. The simulation mentioned above is not simulating DDR3 anyway.

#15

Yes I’m playing with the same board, it’s one of my toy
Good thinking,
I create a rules following the documentation will help a lot…manufacture it and test…
Thanks

#16

Bob Pease was big on building stuff && seeing what happened. He was death on SPICE except (possibly) for a rough idea. He had a statement “Solder is my programming language”. Google him and Jim Williams - they were the two top analog guys in the world. Start here
https://www.edn.com/design/analog/4375963/Slideshow--Remembering-Jim-Williams-and-Bob-Pease

1.21 Gigawatts is spot-on. Build the sucker. Put berg headers on the bus & control lines so you can easily hook up a scope or analyzer. (Remember an analyzer can lie to you - a scope is almost always better. You can get a kick-ass Rygol for \$400).

And never fully trust the datasheet - they can (and sometimes do) lie to you.

Have fun! Keep good notes! Feed and water he family at least once a day (assuming you have one, of course…)

#17

I’m thinking on practical way,
If I limit my impedance from the range 40 to 80 ohms, probably 70 ohms is good don’t you think ?
will it be safe ?

Regardless of the output impedance value of the CPU and input impedance value of DDR3 chip
I see Z CPU = 40 and Z DDR3 = 36, and :

Time = 1/3 * rise time
eR = 4 (approximation for FR4)
C = 11.811 in/nSec (speed of light, in inches per nanosecond)

LR = Time * VP
LR = Time * C / v eR

Thanks

#18

DDR impedance specifications are:
50–60 Ω single-ended
100–120 Ω differential

#19

What does it mean exactly with single ended and differential ?
Example ?

So total Zout+Zin+Track+Via = 60 Ohm for Single ended and
120 Ohm for differential ?

Thanks

#20

No, it is the impedance of the PCB traces themselves, exclusive of any components, with reference to the ground plane on the adjacent layer (or each other for differential).

#21

Single ended = one track in the same layer and
Differential = one track in two layers

?

Thanks

#22

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.