DFI NF4 Ultra-D vdimm mod: come inside.

I realize EMC, the Vref trace showed no huge droop, my point is the datasheet calls for the voltage divider to consist of 100K/100K resistors and for them to NOT be used for other purposes at the same time.

The trick DFI implemented to use a simple divider for Vref generation and then "hey, lets save $.02" using this ALSO as the voltage divider for the Vtt generation is a flawed decision.

I believe the 9173 is amplifying spikes on Vref at the actual divider junction....not smoothed by the caps on Vref near the DIMMs... and the Vref droop is a SUB-HARMONIC of the actual spike, so you need to be looking at a much higher freq. noise spike on Vref directly AT THE JUNCTION of the resistors. That noise spike is not making it over to where you were measuring Vref and was at a higher freq.

Point is, by altering the design to have a 200K pot take the place of the voltage divider input to the 9173....set to 50% point, equivalent to 100K/100K.... which would also let us dial in any desired Vtt "bias" above/below Vmem/2..., and then removing the 2 10ohm resistors. Tie one end of the pot to Vmem, the other end to GND, and the middle pin to the REF input of the 9173. With the REF pin on the 9173 pulled in the air, the Vtt output of the 9173 can be shorted to the REF PAD on the PCboard right next to the Vtt pin of the 9173. This would then have the 9173 drive BOTH Vref and Vtt.

Point being, there will be NO DROOP any longer, Vref wont be subject to the noise/droop of the Vtt line as there will no longer BE ANY noise.

A couple caps should also be added to this Vtt/Vref shorted junction ath the 9173, a 560pf, and a .1uf.

I know the design tweak they made to the 9173 circuit was cute, and saved some $$ and seemed like a stoke of genious, but it is flawed and causing this Vtt droop issue. The 9173 circuit needs to be redesigned BACK to what it should be as correctly instructed in the 9173 datasheet.

Vref doesnt take ALOT of current, it wont over stress the 9173, but it is enough current that there are instantaneous voltage spikes at the 10/10 divider due to current being drawn through the 10ohm resistors and the parasitic inductance the resistors possess. I know you weren't "seeing" it on Vref as measured, but trust me, it is there at a higher freq, and the 9173 is AMPLIFYING it (it being the noise spikes) and so the result coming out of the 9173 on Vtt output is considerably higher than the input goiong into the "ref" sinput of the 9173.

I know you tried a bunch of changes, but try EXACTLY what I suggest here and I'll bet the problem vanishes. Ill go buy another MBoard and try it myself as well.....great like I wanted another motherboard laying around here....I wanted to spend my next MB $ on an integrated video Grouper.

I started to write a long repose, but I'll simply leave it at this: Do yourself a favor bro... don't waste the money on another MB to try it...

As for me... I'll simply wait for my proto PWB to arrive.

Peace :toast:

Quote:

---

Originally Posted by Malves

- -
There is the jumper trick to make more options available in the BIOS. But what if you forget about that and select a vdimm that's higher than your rail? You can kiss your board goodbye
- -

---

Been there, done that. Board and RAM still alive.

Has somebody actually had the board / RAM die by having jumper trick + vDIMM selection higher than the rail?

I had selected 3.6V and had the rail at 3.3V. Nothing weird happened, it booted just like always.

Hi EMC, I was just trying to point out the one and only divergence from the datasheet spec for the 9173 was the Vref voltage divider.

The use of 10/10 ohm actually leads to a "current component" on the REF pin that the 9173 was NOT designed to handle. The design calls for 100K/100K which provides the voltage information for feedback to control the Vout of the 9173, but NO current flow possible to/from the REF pin....since it is 100K to GND and 100K to Vin/ctrl. The difference in potential current flow in the pin (and from the internal 9173 circuit that we dont have the schematic to) between using 10/10 rather than 100K/100K; and the further addition of the fact that the 10/10 is part of an active circuit (Vref supply) is a design element that seriously DIVERGES from the 9173 spec.

Whenever you design a circuit and get an unexpected result/behavior (like the droop spikes on Vtt in this case) you need to first review the circuit for where anything diverges from the spec.

The 9173 is an excellent component, DFI made a great decision using it for this purpose. This droop should not be happening and the reason why it is traces back to where in the circuit does DFI diverge from the spec. I think it clearly is the use of the voltage divider for dual purpose.

The 9173 is amplifying the current spikes that the REF pin is experiencing during operation, and converting it to a voltage spike on the output (Vtt). You wont "see" a corresponding voltage spike on the REF pin because there is no "voltage" spike, but there is an out-of-spec "current" spike at the voltage divider junction which is causing a current flow into/outof the REF pin that it was NOT designed to deal with or expect.... as the spec calls for 100K/100K divider components.

Just wanted you to know Im not smokin anything funny.

uwackme, I'd say go out, and do the mod. If it works, then you'll be the hero! ;)

damn uwackme sounds like you know what your doing i would give your mod a try i dont really see why it could fry the motherboard if what you art trying is incorrect so i would go for it. hehe thats always my attitude about mods tho and i kill alot of hardware :P
________
Tv-Gossip.Com

I agree Cel, I will be doing it. I'm just not in a position to actually measure with a highspeed scope the results. That's why I was mentioning it to EMC.

Quote:

---

Originally Posted by uwackme

I agree Cel, I will be doing it. I'm just not in a position to actually measure with a highspeed scope the results. That's why I was mentioning it to EMC.

---

OK, good to hear. Then all I can still say is good luck and may the overclocking gods be with you. :D
You might not have the equipment, but perhaps the results after modding the board will speak for themselves. Be sure to test the board extensively before modding, so you'll actually see a difference if everything works out fine. After that you could always send the board over to EMC for measuring, if he's willing to.

I think it's one of the mods that everyone is talking about, but noone's got the time, the guts or the skills to actually do it. Just like the 5V Vdimm-source mod on the Socket 754 DFI Lanparty board (OK, I need to stop to praising myself, that's not good :slap: ; to be honest it was pure luck that it worked out fine in the end after all that went wrong :confused2 :wiggle: ).

Have fun and give us a reason to praise you. :) :toast:


BTW, one reason why I'm very interested is that my Lanparty UT NF3 250GB also uses the RT9173 controller. I still have got to measure VTT carefully under load to see, if my board also got these issues (one thing I noticed is that VTT is always a little higher than 1/2 Vdimm on that board, regardless of the set Vdimm; for example at 3.1V Vdimm, VTT is 1.58V), but if it is, IMHO the solution should be very similar to the one on the NF4 Lanparty.

Quote:

---

Originally Posted by uwackme

Just wanted you to know Im not smokin anything funny.

---

:rofl: Just for you, when I do my next round of testing, I'll grab you some results from a seperate divider fed Vref input, as well as a buffered version of a seperate divider fed input. If you want to talk about why I said what I did, PM me and we'll talk ;)

Peace :toast:

Teaser: Problem....solution.

EMC, you wouldnt happen to have a hi-res image of the 9173 area with the two 10ohm resistors of the divider marked?

I want to remove the two 10ohm resistors, pull the REF pin up (supplied by a 200K pot between Vmem and Vgnd, with wiper pin giving Vmem/2 to REF pin) and Ill short the Vref and Vtt signals at the 9173A pads so the 9173 feeds both.

Withoiut question Im positive it is the parasitic current flow at the REF pin on the 9173 that is being amplified and converted to a voltage fluctuation on the Vtt output.

Didn't have a marked up pic... but adding annotation only took a sec ;)

http://img396.imageshack.us/img396/8...ated7qq.th.png

That is between the two pairs of DIMM slots, towards the I/O connectors.

I would suggest you try first without touching the resistors. Just tune your additional divider feeding the input to the RT9173A so it matches the VRef DC level (use a DMM).

Oh... and Oskar has already done the VCtrl = 3.3V mod and posted results back earlier in the thread. It helped some... but didn't eliminate the issue.

Peace :toast:

ps. Never be "positive" until testing a hypothesis... on multiple boards ;)

Quote:

---

Originally Posted by uwackme

EMC, you wouldnt happen to have a hi-res image of the 9173 area with the two 10ohm resistors of the divider marked?

I want to remove the two 10ohm resistors, pull the REF pin up (supplied by a 200K pot between Vmem and Vgnd, with wiper pin giving Vmem/2 to REF pin) and Ill short the Vref and Vtt signals at the 9173A pads so the 9173 feeds both.

Withoiut question Im positive it is the parasitic current flow at the REF pin on the 9173 that is being amplified and converted to a voltage fluctuation on the Vtt output.

---

Any news on this? I have the pot here waiting to reproduce if you have had any joy?

Put away your pot Big Toe...

Too tired from weekend marathon to do a full write-up with pics tonight, but here's a "Brief News Update" :fact:

Unfortunately, my analysis and probable cause back in post 99 (& 121 & 129) appears correct... and Oskar's supposition in post 98 incorrect :(

1) Feeding pin 4 of the RT9173A from a seperate divider makes no difference

2) Feeding pin 4 of the RT9173A with a buffered, extremely filtered reference voltage makes no difference

(sorry wackme... was really hoping I was wrong and one of those would solve it... would have been an easy mod for everyone else)

3) It is possible to eliminate 90%+ of the fluctuation problems... but the largest scale fluctuations can not be eliminated by controlling the RT9173A (at least while using Vmem to source Vctrl) and at this point would say there's a 99%+ probability they are a result of the inability of the RT9173A to source enough current under some conditions (really makes me wonder about the termination circuit on the MB).

4) Two possible solutions - can test both with item in hand ;) One is to supplement the RT9173A, the other to replace it... testing of first in next couple days... second may have to wait till next weekend depending on free time. BTW... the "teaser" is working great so far :D

Full update with pics to follow (maybe tomorrow, depends on work). We now return you to your regularly scheduled programming :p:

Peace :toast:

That's a bummer.

So you think the 9173 is simply being overstressed and internal OC protection is kicking in?

I would have hoped the OC protection would have a hysterisis to it, so the responsiveness wouldnt be there to allow "sort of works" :cool:

That way it would have crapped out on DFI in research and been addressed before production. yuck.

Think piggybacking 2 9137's would solve it?

What value terminator resistors did they use? The NForce2 boards used 68ohm resistors for most signals.

Piggybacking a second 9173A might work... don't have any to try right now.

The termination network on the DFI is a tad more than just resistors (appears they have AC termination as part of the network) and varies by position ;) The pure resistive parts are 39 & 47 ohms...

Write up with annotated pics complete for the part 1 testing... just trying to decide where to post it...

Peace :toast:

Maybe Oskar could give us a snippet from the schematic on the termination network so we can all brainstorm?

Sounds like it is something beyond Jedec DDR or AMD A64 requirements.

It should simply be a "pull-middle" resistor between Vtt and each data/control/address signal to the DIMMs. With alot of bypass caps on the Vtt side of the resistor arrays to make sure Vtt is rock solid on the supply side.

But the huge spikes, since they arent from the regulation control of the 9173 must as you suggest come from it simply being over stressed.

Regarding the termination method - it was done for a reason... higher bus speeds. Just didn't match the Vtt supply to the circuit :slap:

Well... based on intermediate results of "part 2"... no need for the "snippet" ;) Will know more after further testing...

But for now I decided where to post the "part 1" results... in a new thread :p: Don't worry, it'll be easy to recognize :lol:

Peace :toast:

Hi, i have read the whole thread now. Found i wery interesting.. Any news? where exactly did you post the results emc2?

In the first thread in this very section. It's surprising you missed it.

Quote:

---

Originally Posted by craig588

In the first thread in this very section. It's surprising you missed it.

---

not surprising. dont oyu know no one reads stickys? :slap:

Quote:

---

Originally Posted by craig588

In the first thread in this very section. It's surprising you missed it.

---

haha, well im rather surprised too by my own incompetence... :wasntme: sorry guys...

I was reading this thread now and my question is...is there any profict on making the vtt fix that tony did on people that are using tccd ram?

What's the difference on making the vdimm mod with the lm358 with pin4 and 6 and the other method, with the leg of pin 17? Just the ability to turn off?

Cya around ;)

I know I'm late to the party.

To those of you who have boards modded, what do you set voltages to in bios?

Since Vcore can be set waaaay low in bios, i wonder if setting it as low as it can go and then adjusting waaay up with the trimmer is more stable. Has anyone tested?