“ PLL on clock tree was using other, quiet power supply.”
The above is from your original reply.
And you never corrected my statement
" Back to the PLL jitter issue. As long as the clock buffer output is used for
the PLL phase detector input, it is still part of the closed PLL controlled
loop."
The PLL can have completely clean power, as long as the victim clock tree
buffer is driven by a PLL output, the same dirty clock tree will be used to
feed the PLL phase detector that will see the clock tree buffer power noise as
phase noise and try to correct it.
Chris Cheng
Distinguished Technologist , Electrical
Hewlett-Packard Enterprise Company
+1 510 344 4439/ Tel
chris.cheng@xxxxxxx / Email
940 N. McCarthy Blvd., Milpitas, CA 95035
USA
From: Iliya Zamek [mailto:i_zamek@xxxxxxxxx] ;
Sent: Thursday, August 30, 2018 11:07 PM
To: si-list@xxxxxxxxxxxxx; Cheng, Chris <chris.cheng@xxxxxxx>
Subject: Re: [SI-LIST] Re: Controlled ESR capacitor
Actually, I am glad that you wrote this email Chris,
I see that I would better brake this paper into 3-4 with more detail
explanation of each. May be I will do it for the next DesignCon.
I agree with you that PLL significantly impacts jitter, and if we would use
PLL, we should account for it, validate, and choose PLL BW and other
parameters to minimize, or compensate its impact to the measurements. However,
we did not use PLL. In our experiments Victim output inside
FPGA was routed directly to the output I/O for Jitter measurements, and
Aggressor input was routed strait to the input buffer. Where you saw
mentioned PLL in the paper? There is no one place.
Interesting, that this whole research started because PLLs inside FPGA were
blamed for the total high FPGA jitter. I did research, measurements
and was able to demonstrate that PLLs are perfect. We started to dig deeper,
led series of projects for some years, did many measurements.
Have you noted that the resonance peak measured at 5% FPGA Core logic
utilization, only. You would astonished to see what we got at higher
Aggressor logic utilization. It showed us what is going on inside
Microprocessor, ASIC, or FPGA when internal logic is switching - everything is
moving, like on the ship that came to be in the middle of ocean storm. This is
way more stronger noise than from Ozzy and Gene together and I
would not advice you to take an ear muff away. Thank you for the joke.
I am waiting for your further emails.
Best regards,
Iliya
On Thursday, August 30, 2018, 5:20:53 PM PDT, Cheng, Chris
<chris.cheng@xxxxxxx> wrote:
"it was demonstrated that the traditional approach to the reducing PDN
resonances by increasing the decoupling on PCB does not help to remove
resonance peaks."
"Otherwise, lowering the PCB impedance by adding more capacitances makes the
situation even worse, - the values of the peaks increase."
The above statements are the reason why I feel we cannot take this discussion
offline and continue the discussion because these are important concept I
disagree with you. You are making the assumption that jitter response to power
noise is flat across frequency band and the effect of decoupling strategy in
power noise has a proportional impact across all frequency band on jitter. It
does not.
The whole purpose of PLL is to dynamically eliminate phase noise (jitter)
either induced by power or clock source. It can self-correct up to the PLL loop
bandwidth which is set to maintain stability also. The steep rise in jitter
across all configuration between 20-35MHz is not because your power noise is
hitting a resonance, it is because the PLL loop bandwidth is reached and the
PLL starts to lose its capability to correct the jitter. I would even venture
to speculate the 35MHz peaking is due to the PLL loop is slightly underdamped
resulting in jitter peaking.
The point I am trying to say is both high cap and low cap count configurations
have resonance at low frequency, it seems the low cap count happens around
10MHz while the high count around 35MHz. With the PLL loop bandwidth being
above one (low cap count) and below one (high cap count) there are drastic
differences between the two but you can't use that to conclude low cap count
noise is better than the high cap count based on the jitter performance. In
order to do that, you made the assumption power noise to jitter transfer
response is flat between 10 to 35MHz. It is not.
This is like Gene Simmons is playing in one room and Ozzy Osbourne is in
another one. You put an ear muff over my ear and walk me into Gene's room and
then take it off and walk me to Ozzy's room. Then you tell me, "See, Ozzy is
noisy".
Chris Cheng
Distinguished Technologist , Electrical
Hewlett-Packard Enterprise Company
+1 510 344 4439/ Tel
chris.cheng@xxxxxxx / Email
940 N. McCarthy Blvd., Milpitas, CA 95035
USA
From: Iliya Zamek [mailto:i_zamek@xxxxxxxxx] ;
Sent: Wednesday, August 29, 2018 11:49 PM
To: si-list@xxxxxxxxxxxxx; Cheng, Chris <chris.cheng@xxxxxxx>
Subject: Re: [SI-LIST] Re: Controlled ESR capacitor
Hi Chris,
Sorry for the late reply; I was busy after work tonight.
Thank you for reading my paper "On-chip Jitter and system Power Integrity"
DesignCon 2012.
You are still having in mind a design and how we can improve it, but this is
not a case. This paper main goals were:
1) to describe the method for PDN validation through Jitter measurements which
less impacted by parasitic and applicable for any ICs weather they have special
features, or not, that other methods require;
2) using this method we revealed Jitter resonance induced by PDN noise that
origin of switching on-die logic.
Note, that this is a characteristic of PDN performance "looking from the
die side".
3) next, using this method we researched how on-board decoupling impacts the
PDN resonances.
The method: two logic domains in standard FPGA were used that share the same
PDN; one logic domain was used as an Aggressor, and other as a Victim. Both
were switching periodically and asynchronously. Switching frequency
sequentially changing and Jitter of a Victim was each time measured.
Jitter resonance: switching of an aggressor creates pulses of a current
corresponding each rise and fall time of Aggressor signal; these current pulses
causes PDN voltage variations (noise). Changing Aggressor frequency we change
noise frequency; when noise frequency reaches PDN resonance the impact on the
Victim became higher and victim Jitter reaches the peak. So, this is a way to
find a PDN resonances.
Decoupling: we changed on-board decoupling and measured Jitter; we obtained
that with rising decoupling at 36 MHz the resonance became stronger and in
field 100-200 MHz Jitter slightly reduced - this effect also explained in
paragraph following the graph. These measurements are time consuming and we did
not measure higher resonances. I think that Jitter reduction at 100-200 MHz
with increasing decoupling is an evidence that higher frequency resonances also
became more narrow and higher.
So, we did not sacrifice nothing; we just did an experiment, got data, and
explained the effects.
I was glad to see Binayak' good explanation of the effect that I also got from
these experiments: looking from die capacitance through package inductance to
the decoupling we see parallel resonance. Its' ESR is decreasing with rising
number of the capacitances. When ESR is decreased, Q-factor rising and, also,
the resonance peak.
Chris,
I agree with you that these results are important as an illustration of new
effects that has not been noted before: particular, it was demonstrated that
the traditional approach to the reducing PDN resonances by increasing the
decoupling on PCB does not help to remove resonance peaks. Otherwise, lowering
the PCB impedance by adding more capacitances makes the situation even worse, -
the values of the peaks increase.
I will be glad to continue this discussion, which is very interesting for me
personally. However might be not all Si-listers are interested in it; should
we move it off-line?
Best regards,
Iliya
On Wednesday, August 29, 2018, 4:59:50 PM PDT, Cheng, Chris
<chris.cheng@xxxxxxx> wrote:
Your clarification makes me even more puzzled about your conclusions. If this
is general one die core noise to clock tree buffer power, why would you
sacrifice the critical >100MHz noise damping that most core logic will care for
the 30MHz or so that most switching logic don't care ? Is your chip only
working in 30-40MHz ?
Back to the PLL jitter issue. As long as the clock buffer output is used for
the PLL phase detector input, it is still part of the closed PLL controlled
loop. You can see below 20MHz the PLL loop bandwidth corrected the induced
jitter on the clock tree buffer. You just luck out that when the reduced cap
configuration peaked around 10MHz, that was below the PLL loop bandwidth. Had
the PLL loop bandwidth is forced to be lowered by crazy PLL spec like PCIE to
5MHz, the 10MHz peaking of the red curve will most probably be higher than the
high cap count configuration peak around 35MHz.
Sorry, these are important design trade-offs that I am not sure you have drawn
the right conclusion. We should keep this discussion going.
Chris Cheng
Distinguished Technologist , Electrical
Hewlett-Packard Enterprise Company
+1 510 344 4439/ Tel
chris.cheng@xxxxxxx / Email
940 N. McCarthy Blvd., Milpitas, CA 95035
USA
From: Iliya Zamek [mailto:i_zamek@xxxxxxxxx]
Sent: Wednesday, August 29, 2018 1:28 PM
To: Cheng, Chris <chris.cheng@xxxxxxx>
Cc: dmarc-noreply@xxxxxxxxxxxxx; binayaks@xxxxxxxxx <binayaks@xxxxxxxxx>;
tom@xxxxxxxxxxxxxxxxx; si-list@xxxxxxxxxxxxx <si-list@xxxxxxxxxxxxx>; FRED
MOTTER <motter@xxxxxxxxxxxxx>; FRED B <FRED@xxxxxxxxxxxxx>; Jacques Tazartes
<jacquest15@xxxxxxxxxxxxx>; EmadIbrahim <EIbrahim@xxxxxxxxx>;
Myrna5@xxxxxxxxxxxxxx <Myrna5@xxxxxxxxxxxxxx>; Istvan Novak
<istvan.novak@xxxxxxxxxxx>; junichi.yamada.ak@xxxxxxxxxxx
<junichi.yamada.ak@xxxxxxxxxxx>; Cosmin Iorga <ci42775@xxxxxxxxx>
Subject: Re: [SI-LIST] Re: Controlled ESR capacitor
Hi Chris,
Thank you for your comments.
Both, 30-40 and 100-200 MHz jitter
components are important. There is no
point to sacrifice one, or other.
This was special experiment to validate
the role of board decoupling on
interaction of two logic domains inside
the chip which shared the same
power supply: core logic was switching
(Aggressor) impacting routing
component's and clock tree' jitter (victims).
PLL on clock tree was using other, quiet
power supply. Scale is the same as on
figure 4: (jitter at core switching)/(jitter
at core quiet).
Let me know if you have more questions.
Thank you.
Iliya
On Aug 29, 2018 11:10 AM, "Cheng, Chris" <chris.cheng@xxxxxxx> wrote:
Why do you think it is worth sacrificing the 100-200MHz jitter performance with
high cap count for the narrow range of 30-40MHz jitter performance with lower
cap count ?
In the world of XbitY coding for the 10+Gb/s world, I would think the
100-200Mhz jitter performance is much more important than those 30-40MHz.
To me, the best conclusion I can draw from the two graph seems to point out the
price we pay for the obsession of jitter transfer over jitter accumulation.
Clearly the 30-40Mhz jitter peaking could be fixed by uping the PLL loop
bandwidth from around 10MHz to 40+MHz. In the name of saving a few $ in cheap
clock source, we paid this price for all PLL jitter accumulation downstream.
To recap :
a) Who would feed a PLL with raw unfiltered power and rely on decoupling cap
ESR to maintain PLL jitter performance ?
b) Why is it worth sacrificing 100-200MHz jitter performance to a lower band
30-40MHz ?
Also, what was the vertical scale of the graphs ?
Chris Cheng
Distinguished Technologist , Electrical
Hewlett-Packard Enterprise Company
+1 510 344 4439/ Tel
chris.cheng@xxxxxxx / Email
940 N. McCarthy Blvd., Milpitas, CA 95035
USA
-----Original Message-----
From: si-list-bounce@xxxxxxxxxxxxx [mailto:si-list-bounce@xxxxxxxxxxxxx] On ;
Behalf Of Iliya Zamek
Sent: Monday, August 27, 2018 8:39 PM
To: binayaks@xxxxxxxxx
Cc: dmarc-noreply@xxxxxxxxxxxxx; Cheng, Chris <chris.cheng@xxxxxxx>;
tom@xxxxxxxxxxxxxxxxx; si-list@xxxxxxxxxxxxx; FRED MOTTER
<motter@xxxxxxxxxxxxx>; FRED B <FRED@xxxxxxxxxxxxx>; Jacques Tazartes
<jacquest15@xxxxxxxxxxxxx>; EmadIbrahim <EIbrahim@xxxxxxxxx>;
Myrna5@xxxxxxxxxxxxxx; Istvan Novak <istvan.novak@xxxxxxxxxxx>;
junichi.yamada.ak@xxxxxxxxxxx; Cosmin Iorga <ci42775@xxxxxxxxx>
Subject: [SI-LIST] Re: Controlled ESR capacitor
Somehow my email was distorted. Re-sending.
Hi Binayak,
Your effect description is correct.
If you look at papers which links were included in my email sent 2 days ago
(August 25) you can see the graph that described exactly the same - how low
PCB impedance causes rising on-die transients.
If you could not find these papers, you might connect me off-line and I'll send
it to you.
Thank you.
Iliya.
intensive on-board decoupling does not reduce PDN resonances respect to the
chip die, but even make PDN performance worse, - resonances are rising.
I will be glad to continue this discussion, however might be not all Si-listers
are interested in it and we should move it off-line?
Best regards,
Iliya
On Wednesday, August 29, 2018, 4:59:50 PM PDT, Cheng, Chris
<chris.cheng@xxxxxxx> wrote:
Your clarification makes me even more puzzled about your conclusions. If this
is general one die core noise to clock tree buffer power, why would you
sacrifice the critical >100MHz noise damping that most core logic will care for
the 30MHz or so that most switching logic don't care ? Is your chip only
working in 30-40MHz ?
Back to the PLL jitter issue. As long as the clock buffer output is used for
the PLL phase detector input, it is still part of the closed PLL controlled
loop. You can see below 20MHz the PLL loop bandwidth corrected the induced
jitter on the clock tree buffer. You just luck out that when the reduced cap
configuration peaked around 10MHz, that was below the PLL loop bandwidth. Had
the PLL loop bandwidth is forced to be lowered by crazy PLL spec like PCIE to
5MHz, the 10MHz peaking of the red curve will most probably be higher than the
high cap count configuration peak around 35MHz.
Sorry, these are important design trade-offs that I am not sure you have drawn
the right conclusion. We should keep this discussion going.
Chris Cheng
Distinguished Technologist , Electrical
Hewlett-Packard Enterprise Company
+1 510 344 4439/ Tel
chris.cheng@xxxxxxx / Email
940 N. McCarthy Blvd., Milpitas, CA 95035
USA
From: Iliya Zamek [mailto:i_zamek@xxxxxxxxx] ;
Sent: Wednesday, August 29, 2018 1:28 PM
To: Cheng, Chris <chris.cheng@xxxxxxx>
Cc: dmarc-noreply@xxxxxxxxxxxxx; binayaks@xxxxxxxxx <binayaks@xxxxxxxxx>;
tom@xxxxxxxxxxxxxxxxx; si-list@xxxxxxxxxxxxx <si-list@xxxxxxxxxxxxx>; FRED
MOTTER <motter@xxxxxxxxxxxxx>; FRED B <FRED@xxxxxxxxxxxxx>; Jacques Tazartes
<jacquest15@xxxxxxxxxxxxx>; EmadIbrahim <EIbrahim@xxxxxxxxx>;
Myrna5@xxxxxxxxxxxxxx <Myrna5@xxxxxxxxxxxxxx>; Istvan Novak
<istvan.novak@xxxxxxxxxxx>; junichi.yamada.ak@xxxxxxxxxxx
<junichi.yamada.ak@xxxxxxxxxxx>; Cosmin Iorga <ci42775@xxxxxxxxx>
Subject: Re: [SI-LIST] Re: Controlled ESR capacitor
Hi Chris,
Thank you for your comments.
Both, 30-40 and 100-200 MHz jitter
components are important. There is no
point to sacrifice one, or other.
This was special experiment to validate
the role of board decoupling on
interaction of two logic domains inside
the chip which shared the same
power supply: core logic was switching
(Aggressor) impacting routing
component's and clock tree' jitter (victims).
PLL on clock tree was using other, quiet
power supply. Scale is the same as on
figure 4: (jitter at core switching)/(jitter
at core quiet).
Let me know if you have more questions.
Thank you.
Iliya
On Aug 29, 2018 11:10 AM, "Cheng, Chris" <chris.cheng@xxxxxxx> wrote:
Why do you think it is worth sacrificing the 100-200MHz jitter performance with
high cap count for the narrow range of 30-40MHz jitter performance with lower
cap count ?
In the world of XbitY coding for the 10+Gb/s world, I would think the
100-200Mhz jitter performance is much more important than those 30-40MHz.
To me, the best conclusion I can draw from the two graph seems to point out the
price we pay for the obsession of jitter transfer over jitter accumulation.
Clearly the 30-40Mhz jitter peaking could be fixed by uping the PLL loop
bandwidth from around 10MHz to 40+MHz. In the name of saving a few $ in cheap
clock source, we paid this price for all PLL jitter accumulation downstream.
To recap :
a) Who would feed a PLL with raw unfiltered power and rely on decoupling cap
ESR to maintain PLL jitter performance ?
b) Why is it worth sacrificing 100-200MHz jitter performance to a lower band
30-40MHz ?
Also, what was the vertical scale of the graphs ?
Chris Cheng
Distinguished Technologist , Electrical
Hewlett-Packard Enterprise Company
+1 510 344 4439/ Tel
chris.cheng@xxxxxxx / Email
940 N. McCarthy Blvd., Milpitas, CA 95035
USA
-----Original Message-----
From: si-list-bounce@xxxxxxxxxxxxx [mailto:si-list-bounce@xxxxxxxxxxxxx] On ;
Behalf Of Iliya Zamek
Sent: Monday, August 27, 2018 8:39 PM
To: binayaks@xxxxxxxxx
Cc: dmarc-noreply@xxxxxxxxxxxxx; Cheng, Chris <chris.cheng@xxxxxxx>;
tom@xxxxxxxxxxxxxxxxx; si-list@xxxxxxxxxxxxx; FRED MOTTER
<motter@xxxxxxxxxxxxx>; FRED B <FRED@xxxxxxxxxxxxx>; Jacques Tazartes
<jacquest15@xxxxxxxxxxxxx>; EmadIbrahim <EIbrahim@xxxxxxxxx>;
Myrna5@xxxxxxxxxxxxxx; Istvan Novak <istvan.novak@xxxxxxxxxxx>;
junichi.yamada.ak@xxxxxxxxxxx; Cosmin Iorga <ci42775@xxxxxxxxx>
Subject: [SI-LIST] Re: Controlled ESR capacitor
Somehow my email was distorted. Re-sending.
Hi Binayak,
Your effect description is correct.
If you look at papers which links were included in my email sent 2 days ago
(August 25) you can see the graph that described exactly the same - how low
PCB impedance causes rising on-die transients.
If you could not find these papers, you might connect me off-line and I'll send
it to you.
Thank you.
Iliya.
------------------------------------------------------------------
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 forum is accessible at:
http://tech.groups.yahoo.com/group/si-list
List archives are viewable at:
//www.freelists.org/archives/si-list
Old (prior to June 6, 2001) list archives are viewable at:
http://www.qsl.net/wb6tpu
------------------------------------------------------------------
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 forum is accessible at:
http://tech.groups.yahoo.com/group/si-list
List archives are viewable at:
//www.freelists.org/archives/si-list
Old (prior to June 6, 2001) list archives are viewable at:
http://www.qsl.net/wb6tpu
------------------------------------------------------------------
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 forum is accessible at:
http://tech.groups.yahoo.com/group/si-list
List archives are viewable at:
//www.freelists.org/archives/si-list
Old (prior to June 6, 2001) list archives are viewable at:
http://www.qsl.net/wb6tpu