[SI-LIST] Re: DDR Vref Bypassing
- From: Bill Wurst <billw@xxxxxxxxxxx>
- To: si-list@xxxxxxxxxxxxx
- Date: Thu, 21 Jul 2005 13:08:25 -0400
Steve,
I believe we are perhaps analyzing different physical structures.
It seems that the question is whether a single filter capacitor from
Vref to ground (low pass network) is better than a two capacitor divider
(all pass network with an attenuation factor of 2).
Let's start with the Micron DDR2 DIMM spec, which states,
"VREF is expected to equal VDDQ/2 of the transmitting device
and to track variations in the DC level of the same.
Peak-to-peak noise (non-common mode) on VREF may not exceed
±1 percent of the DC value. Peak-to-peak AC noise on VREF may
not exceed ±2 percent of VREF (DC). This measurement is to be
taken at the nearest VREF bypass capacitor."
It also states,
"Data timing is now referenced to VREF, provided the
DQS slew rate is not less than 1.0V/ns. If the DQS slew rate
is less than 1.0V/ns, then data timing is now referenced to
VIH(AC) for a rising DQS and VIL(DC) for a falling DQS."
Either configuration (low-pass or all-pass) can be made to work. A
resistor/capacitor (x2y) divider at the transmitting device (all-pass),
carried on a trace along with the data to the Vref pins of the receiving
device can give slightly better timing margins (up to 18ps at the worst
case of 2% pk-pk noise and 1V/ns slew rate) as follows. The 1% matching
of the x2y capacitor will keep non-common mode noise to less than 0.5%
of the dc value. As you previously stated, copious decoupling of
Vdd/Vss is required to keep pk-pk ac noise within 2%, and this amount of
decoupling is perhaps twice better than minimum good design practice.
In this way, both the data and Vref signals experience almost identical
Vdd/Vss noise all the way to the receiving device. The cancellation is
not perfect because a logic '1' only experiences Vdd noise while a logic
'0' only experiences Vss noise; each is attenuated by roughly a factor
of two due to both end terminations. Vref is the difference of Vdd and
Vss, divided by two, so the two approximately cancel if, and only if,
the noise on each rail is approximately equal and in-phase. If the
noise is unequal or out-of-phase, then the cancellation is partial, or
worse. But is it worse than the low-pass scheme, which takes an average
of all situations? I believe to answer this question will require
detailed simulation of memory, PCB, controller, and PDS.
Unfortunately, there is application literature out there suggesting both
techniques. Which is better - perhaps we are trying to split hairs.
Regards,
-Bill
/************************************
/ billw@xxxxxxxxxxx /
/ /
/ Advanced Electronic Concepts, LLC /
/ www.aec-lab.com /
************************************
===========================================================
steve weir wrote:
> At 09:01 PM 7/19/2005 -0400, Bill Wurst wrote:
>
>>Steve,
>>
>>I'll try to answer your comments/questions as I understand them (see
>>response below).
>>
>>Regards,
>>
>> -Bill
>>
>>
>>======================================
>>steve weir wrote:
>>
>>>Bill, I really question where this assumption of average between Vdd and
>>>Gnd comes from. The noise level on Vdd at the transmitter is not going to
>>>be the same as the instantaneous noise at the receiver, so I don't buy an
>>>argument based on the launch. That leaves us with the transmission
>>>channel. What percentage of what couples onto the transmitted signal
>>>depends on how the board is constructed.
>>
>>If the noise on Vdd and Vss at the transmitter propagate to the receiver
>>with the same velocity as the signal, then the noise at both the
>>transmitter and receiver will be the same but delayed by the flight
>>time.
>
>
> But I think it doesn't. If I have decent power bypass, then the radial
> disturbance on the planes encounters bypass for the transmitter first,
> where most is supposed to reflect, ( or else we don't have a very good
> bypass network ), and a little moves on outward, including onward towards
> the receiver. When that wave hits the receiver's bypass network, it again
> reflects, and if we have done a job well, very little is left to get to the
> receiver. Conversely, the bypass network near the receiver localizes the
> receiver's current noise.
>
> So, I conclude that by coupling Vdd into Vref I conclude we are getting:
>
> Vref_ac = vector sum ( Vnoise_vss_rec * K1 + Vnoise_vdd_rec * K2 +
> Vnoise_vss_tx * K3 + Vnoise_vdd_tx * K4 )
>
> Now, I subtract that from my signal that is:
>
> Vsig_ac = vector sum ( Vterm * K5 +/- ( Vdd_tx - Vss_tx ) * K6 +
> Vcavity_ac * K7 )
>
> Since K6 is ideally close to 1.0, K3 and K4 are ideally << 1, and K1 and
> K2 are close to 0.5, I see essentially no cancellation of the ( Vdd_tx +
> Vss_tx )/2, and injection of a lot of Vnoise_vdd_rec.
>
> Where have I gone wrong?
>
>
>>Granted, the construction of the board can and will impact this
>>assumption. The premise is to make the noise common to both sides of
>>the differential receiver, so that the receiver will reject the noise by
>>virtue of its CMRR.
>
>
> Agreed on premise, disagree mightily on implementation.
>
>
>
>>At launch, a logic '1' will replicate Vdd noise
>>which will be attenuated by terminations at both ends and also
>>influenced by any noise on the reference planes as it travels to the
>>receiver. Similarly, a logic '0' will replicate Vss noise. Trying to
>>make Vref equal to 50% of the difference between the two rails is
>>admittedly a compromise and, as you point out, other factors will reduce
>>the cancellation further. Yet I fail to see a better alternative.
>
>
> Please see above.
>
>
>>>A split filter is in essence a 2Y common mode filter turned inside
>>>out. Impedance mismatch gives rise to mode conversion.which throws off
>>>that 50% divider assumption for equal value ( data sheet )
>>>capacitors. This is why we see 2Y RFI filters with a much bigger X
>>>capacitor shunting the two lines together- to swamp out the mode
>>>conversion. In the Vref application, the X capacitor is represented by
>>
>>the
>>
>>>bypass network from Vcc to Vss. That is really ugly, because it basically
>>>says that we need to bypass the heck out of Vdd to get around mode
>>>conversion in the Vref bypass caps.
>>
>>I agree. This is essentially a 2Y CM filter, and Vdd must be bypassed
>>to the greatest extent practical.
>>
>>>The two capacitors in an X2Y match so well that even for analog
>>>instrumentation they do not need an X capacitor. I have an application
>>>note on this in ADI's instrumentation amplifier designer's guide, based on
>>>real circuit measurements. An X2Y configured as: Terminal A => Vdd,
>>>Terminal B => Vss, Terminals G1, and G2 => Vref matches to better than
>>>1%. So, if one is bent on implementing the divider, X2Y capacitors do the
>>>job in a way that is basically impossible using separate capacitors to
>>
>>each
>>
>>>rail.
>>
>>Again, I agree. I hedged in my response because I was unsure of the
>>matching that could be achieved with x2y capacitors.
>>
>>>If someone is really bent on this divider approach, then X2Y is definitely
>>>the way to go. But given that people have been building with it, and
>>>apparently it has "worked" despite the mode conversion with regular
>>
>>caps, I
>>
>>>really question the validity of the approach in the first place. Do you
>>>know what the physical basis for the rationale of the divider is supposed
>>>to be?
>>
>>Hopefully, unless I've missed something, I've answered this in my
>>response to the first paragraph. Please let me know if I haven't.
>>
>>>Regards,
>>>
>>>
>>>Steve.
>>>
>>>At 12:48 PM 7/19/2005 -0400, Bill Wurst wrote:
>>>
>>>
>>>>Chris,
>>>>
>>>>The answer to your question lies in understanding the function of the
>>>>Vref line. DDR, as well as DDR2, utilize differential receivers to
>>>>process single-ended inputs that have been generated by drivers which
>>>>swing in a balanced fashion around the mid-point of the VDD/GND system.
>>>> To properly process these single-ended inputs, the inverting input of
>>>>each differential receiver is connected to Vref. The receivers will
>>>>work best when Vref equals exactly 0.5*(VDD-GND), including any noise
>>>>that is present on the VDD/GND system. The purpose of placing an equal
>>>>amount of capacitance from Vref to VDD and from Vref to GND is to form
>>>>an ac divider that keeps Vref equal to 0.5*(VDD-GND) over all
>>>>frequencies. The capacitance should be large enough to swamp out any
>>>>parasitic capacitance that exists which could imbalance Vref.
>>>>
>>>>I'll have to hedge on the second question which was "whether an x2y
>>>>capacitor is better than two discrete capacitors" since I don't know
>>>>enough about x2y devices. Properly configured, an x2y capacitor could
>>>>perform better, but the bottom line comes down to the accuracy of the ac
>>>>divider.
>>>>
>>>>Regards,
>>>>
>>>> -Bill
>>>>
>>>>
>>>> /************************************
>>>> / billw@xxxxxxxxxxx /
>>>> / /
>>>> / Advanced Electronic Concepts, LLC /
>>>> / www.aec-lab.com /
>>>> ************************************
>>>>=================================================================
>>>>Christopher R. Johnson wrote:
>>>>
>>>>
>>>>>I have seen references that have Vref bypass capacitors to both VDD and
>>>>>GND. Other references have capacitors only to GND. Is it really
>>>>>necessary to have "balanced" capacitors on the Vref lines? Why? If the
>>>>>"balanced" design is desirable, would an X2Y capacitor be a good choice,
>>>>>since it is "2 capacitors in one"?
>>>>>
>>>>>Regards,
>>>>>
>>>>>Chris Johnson
>>>>>------------------------------------------------------------------
------------------------------------------------------------------
To unsubscribe from si-list:
si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field
or to administer your membership from a web page, go to:
//www.freelists.org/webpage/si-list
For help:
si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field
List FAQ wiki page is located at:
http://si-list.org/wiki/wiki.pl?Si-List_FAQ
List technical documents are available at:
http://www.si-list.org
List archives are viewable at:
//www.freelists.org/archives/si-list
or at our remote archives:
http://groups.yahoo.com/group/si-list/messages
Old (prior to June 6, 2001) list archives are viewable at:
http://www.qsl.net/wb6tpu
Other related posts:
