Vtt on DFI nF4 versus EMC2 on Java :-P

DFI, Vtt, and Me
More Excerpts from a Java Junky

Once again I decided to step into the ring, this time with the opponent being Vtt on the DFI nF4 :lsfight:

As shown in previous scope captures in this thread , Vtt fluctuates out of spec on the DFI nF4 under some conditions... in some cases badly, weaving like a college frat brother after a New Year's Eve bash :lol:

There have been a number of suppositions and theories put forth as to possible causes and fixes, and I finally had time to pull out ye ole Java beans, jumper wires, and start the proceedings :fight: <-that's me on the right :p: Decided to do this one as a bit of a doc-u-modder-ry :hehe:

First up was getting a good "lay of the land" to prevent releasing the magic smoke :smoke: by probing around on the MB :poke: .

Typical of this stage was looking at Vref and Vmem at the DIMMs:
http://img395.imageshack.us/img395/7...ms132xy.th.png

After coming up with a plan of attack, work commenced on a first stage proto circuit. The usual workplace is full, so I had to clear off a new area to work elsewhere. Here's a view of the temporary modder's "nest":
http://img215.imageshack.us/img215/8...o1173ey.th.png
My old Amiga's mouse decided to watch the proceedings :lol: And here's a closeup of the first test widget to be cobbled together:
http://img215.imageshack.us/img215/2...dy149hd.th.png

After initial testing and probing of the enemy's weaknesses :weapon: , decided a better weapon was needed for this match.

Pulled out the drawing board and came up with a new weapon :idea: to combat the evil MB :hrhr: Ordered new supplies for the battle and then came the part I hate...waiting :brick: Kind feel like Wiley Coyote waiting on Acme to deliver the latest goodies, checking for deliveries each day after work :lol:

The first box of supplies arrives:
http://img395.imageshack.us/img395/2...rs117lt.th.png

And after what seemed like forever, the last piece of the new weapon arrives on a Thursday:
http://img399.imageshack.us/img399/5...wb150kw.th.png

Only one day to go, then it's the weekend :wiggle: Then it's time to get out the java :coffee:, put on some tunes :dj:, and start assembly :hammer:

Initial components on the back side of the board:
http://img142.imageshack.us/img142/4...kpb10lj.th.png

Stage 1 circuit complete, perspective of the size of the board:
http://img395.imageshack.us/img395/9...nd126ol.th.png
Doesn't look like much, does it? Well... let's just say looks can be deceiving :hehe:

Now to prep some test leads for the board. Need ends that can be clamped in the wire blocks easily and not get frayed, so we do this:

Strip and form a loop:
http://img398.imageshack.us/img398/9...1a208xk.th.png

Tin the wire to give it rigidity and keep it from fraying:
http://img398.imageshack.us/img398/2...1b215zl.th.png

Slide and shrink the insulation over it:
http://img398.imageshack.us/img398/2...1c221hn.th.png

And now for a little test fit to make sure didn't make it too big:
http://img215.imageshack.us/img215/6...e1183dx.th.png

Needed a test lead with a built in jumper too, so use an old trick like this...

Prepping:
http://img398.imageshack.us/img398/4...2a239yx.th.png

Finished ends:
http://img398.imageshack.us/img398/9...p2b15jt.th.png

Of course, we always do a little bench testing first ;)

Here's feeding the little fella with some voltage:
http://img399.imageshack.us/img399/2...put31xu.th.png

And here's a quick check of the output side of things:
http://img399.imageshack.us/img399/4...put41za.th.png
Looking good so far :)

Now to get the MB ready for insertion... first item up is lifting the leg on the RT9173A:
http://img395.imageshack.us/img395/7...17392lj.th.png

Now we have to put a test lead wire on the pad:
http://img363.imageshack.us/img363/5...d4196zs.th.png

Don't want to short anything out later, so next we insulate our work:
http://img215.imageshack.us/img215/4...d4162bc.th.png

Now we're almost ready to commence round two after a lot of building and prep work... but first just like Wiley Coyote, we have to break for some lunch...

Prepping a quick bite to eat :slobber:
http://img395.imageshack.us/img395/7...me104zh.th.png
Didn't think I survived on Java alone, did ya? :p:

First up, we connect the jumper to the lifted leg and make sure we didn't kill anything during the mod to the MB:
http://img363.imageshack.us/img363/6...rks80yx.th.png
Notice the red jumper wire where the RT9173A resides (overlook the mess in these pics... in the middle of getting ready to move)

MB still ok, now to hook up the new toy with power from the MB and see if we're still ok:
http://img363.imageshack.us/img363/9...red72za.th.png
No magic smoke emminating from anything... still looking good :)

Time to put the camera away and get out the scope (no, not for my breath you smart alecks... one like this -> :scope: )

First thing is to see how well the filtering works on the VRef input from the MB:
http://img215.imageshack.us/img215/3...t3279pl.th.png
Looking good... all the base MB frequencies are squashed that were so predominate on previous captures at the memory and bus clocks rates. Noise used to be 146mV peak-to-peak... now it's under 20mV and "out-of-band".

Now for a look at the noise on the output of the circuit:
http://img215.imageshack.us/img215/2...t1284cr.th.png
The result is overlayed on top of the input noise analysis. The main noise component at 730Mhz is down another 15dB...now the peak noise is up at almost 900Mhz, but about 73dB down. Looking very good, less than 10mV peak-to-peak noise :D And only have half the high frequency caps are on the board ;) As a point of comparison, the standard fluctuations on most Vtt circuits is up to 4 times the total noise seen here :lol:

Ok... now to start checking out some things. Decided to use MemTest 8 for the test condition. It doesn't give the worst fluctuations I've seen, but it does give repeatable results. An important note of sorts about the following scope pics... the Vref signal is recorded with full bandwidth of the test equipment... better than 1Ghz. The Vtt signal is bandwidth limited to 20Mhz so the DC changes can be seen (Vtt looks downright fugly without the limiting :sick: )

Here's one that will be familiar to anyone that's followed this saga, Vtt during T8 with the MB in it's stock condition:
http://img363.imageshack.us/img363/1...ndard363sv.png

I told Uwackme I would investigate his idea and take some pics with the Vref input to the RT9173A connected to a seperate divider from the MB, so those are the first tests.

Here's Vtt during T8 with VrefIn (pin 4) of the RT9173A connected to a seperate divider connected to filtered power (configuration 1):
http://img363.imageshack.us/img363/6...vider358hz.png
As you can see, Vtt on the MB continues it's same old song and dance :(

Here's test 2's results. This time VrefIn is being driven by a buffered version of the seperate divider to completely isolate it (configuration 2):
http://img363.imageshack.us/img363/9...vider344wh.png
Still no joy :( The bell rings, the goils walk the floor :cheer: and then we commence...

* ROUND 2 - First Stage Enabled * :lsfight:

Now it's time to engage the first stage of the new weapon and see what happens...

A capture during the initial tests in Memtest while verifying the MB still worked:
http://img363.imageshack.us/img363/3...itial308xa.png
You'll notice the Vref input to the RT9173A is being modulated now to force it to properly maintain Vtt :woot:

Sat and watched Memtest running through it's paces and grinning whilst Vtt remained rock steady. Scope setup to trigger on any fluctuation greater than 10mV and no triggers :D Then we get up to T8...

... and we get mixed results. The first stage circuit works great, and Vtt remains steady except for in one spot. Spend some time fiddling and probing and end up changing the voltage scale on the Vref input to the RT9173A so I can see it's max swing. The results a pretty damning evidence pointing to the RT9173A being simply unable to supply the current demand being placed on it by the Vtt termination network on the MB. Check it out, T8 again:
http://img363.imageshack.us/img363/3...imits312mi.png
Notice all the fluctuations are gone except for the one set of dips. Stage 1 is driving the input reference to the RT9173A all the way up to almost 2.3V (this with nominal value around 1.4V) and the RT9173A is simply refusing (unable) to put out anymore current to the Vtt termination :( I can hear the poor little thing saying, "I'm given her all she's got Scotty!" :lol: Definitely looks like DFI missed the boat here :slapass:

Here's another capture of running T8 with the first stage engaged, with the voltage scale on the VrefIn to the RT9173A adjusted down so we can see how it's being modulated to maintain the steady Vtt as long as the RT9173A is able to handle the current load:
http://img363.imageshack.us/img363/1...tsmid322lc.png
You can see how the VRefIn to the RT9173A modulates to maintain the nice steady Vtt between the two overload conditions. Can also notice there is one other point that the load current is getting up there (the point just above the 'M' in EMC2).

In conclusion and to quote a famous phrase, "Houston, we have a problem." There exists load conditions that the Vtt supply circuit on the MB simply can not handle.

The good news (of sorts) is the first stage of the new "weapon" works superbly, and can correct most Vtt fluctuation issues on MBs and the Vtt issue is much more clearly defined in the case of the DFI nF4, which is two-fold. The first, fluctuations due to load changes of a moderate level is easily correctable. The second, of those fluctuations due to apparent overloading of the RT9173A will take a bit more to overcome.

Oh... notice I've been saying "first stage"... well, there's a reason... I never come to a serious :lsfight: without a proper light saber of swiss army knife quality :hrhr:

First, we have "configuration 4" of the lightsaber to test using "stage 1"... which will allow supplementing the output of the RT9173A. And if that isn't successful... well, there's always "stage 2" :frag: to enable :wasntme:

So concludes another episode of the chronicles of a Looney Java Junky :p:

Peace :toast:

Preliminary Results of Round 3 :lsfight:
He's baaaaaa-aaaaaaaack :p:

Well, work has conspired against me getting much test time in this week and testing is not finished yet, but thought I would share some of the preliminary results of "Stage 1/Config 4". :)

This time instead of controlling the RT9173A's output via it's Vref In signal, the "new weapon" was configured to strictly augment the Vtt rail in parallel to the RT9173A.

Well, they say a pic is worth a thousand words, so here is what happened to Vtt when hit with the augmentation :hehe:

The first few minutes of the round:
http://img138.imageshack.us/img138/9...cope4mu.th.jpg
Needless to say, I looked something like this when I saw that -> :woot:

Late last night I saved and annotated this towards the end of a long test run so one can see the results clearly:
http://img386.imageshack.us/img386/5...dstage14cu.png
:D

One note here...I'm using long, small Awg jumpers and SMD clip-ends with low amp ratings for this initial testing ... hence the Vmem input to the PWB showing some droop. Makes for some nice "outside-the-specs" testing conditions though :hehe:


Now while pics of Vtt are good... there is always testing of other sorts ;) Before I started making use of the little board that could, I did some baseline testing for comparison to know what affect a steady Vtt might have on things. I wanted to isolate the memory as best could, so all testing is done with a 8x CPU multiplier.

BASELINE TESTS
First up, a nice long baseline using T5 :
http://img386.imageshack.us/img386/4...pass7lf.th.jpg

Not bad, 292Mhz for T5... although I've run higher with a different CPU in the past.

Now we switch over to T8... first at the same 292Mhz:
http://img138.imageshack.us/img138/1...fail1ba.th.jpg
Nope, fails on the 4th pass :hm:

Ok,what about 288Mhz?
http://img138.imageshack.us/img138/3...fail7iv.th.jpg
Well, it made it to the 5th pass :lol:

Ok, let's try 284Mhz:
http://img138.imageshack.us/img138/2...pass3bo.th.jpg
A little 14 Pass, 33 minute run with no errors... 8Mhz below what T5 ran at.


TEST RESULTS USING Vtt AUGMENTATION

Here are some of the results so far with the vastly improved Vtt shown in the scope capture above with the little board that could:

First up, a view of some late night testing this week with the improved Vtt:
http://img135.imageshack.us/img135/4...view1em.th.jpg

Was going to originally show the closeup pic of that run for this post, but when I was assembling and annotating the pics, I noticed it wasn't as long of a run as the baseline, so....

....a 24 pass, 53 minute run of T8 - longer than the baseline T8 test and most benchies (sorry, forgot to turn the flash off :hitself: ...something about sleep deprivation and celebration :p: )
http://img138.imageshack.us/img138/2...nted4eg.th.jpg
Yes, that's right... 14Mhz higher than the baseline... and 6Mhz above the T5 run :banana: Quite obviously, Vtt does make a difference :p:


NOTE: ALL test results shown were done with the EXACT same BIOS settings (except for FSB), same memory, CPU, etc., etc. The ONLY diffs are FSB speed and Vtt configurations.


On a silly side note, I think that's the first :banana: I've used on here :rofl:

Well, it's a short update... but it sure was sweet :D

--------- Update to post, Round 3 TCCD roundup ----------

A few updates to this part of things...

First, this is where the TCCD ended up at with long runs of MemTest:
http://img206.imageshack.us/img206/2...4h4m3mi.th.jpghttp://img206.imageshack.us/img206/7...4min0nd.th.jpg
MemTest T8 ended up @ 297Mhz (298Mhz had an error every 80 mins or so)... MemTest T5 showed a small gain of 2Mhz to 294Mhz from the baseline (Vtt only has one smaller fluctuation at the end of it, so no surprise there). Important to note once again, there are NO changes to BIOS settings other than MemClock speeds and NO hardware changes. Who knows... maybe it could now be tweaked with a bit tighter timings or up another couple Mhz with different drive strengths... time allowing I might just check ;)


And here are some further shots of Vtt for comparison:

In this one I switched off the circuit half way thru a scope capture:
http://img206.imageshack.us/img206/7...isabled2wr.png

And here is where I switched it back on:
http://img380.imageshack.us/img380/5...enabled9in.png

You can see both stock DFI and augmented DFI Vtt on the same pics now ;)


And here are pics showing the positive and negative Vtt fluctuations with the augmented Vtt... that 8mV fluctuation noted in the first part of this post is actually a +/-4mV fluctuation from nominal...
http://img380.imageshack.us/img380/6...dmaxneg7vn.png
http://img132.imageshack.us/img132/9...dmaxpos0lp.png


so we ended up going from a 72mV droop to a 4mV droop. Not bad, eh? :D Oh... one other small comment, if you look in the original thread, those Vtt fluctuations can get much worse without augmentation....

* Minor Update, Stability Overnight *
And here's a pic of the overnight run off T1 thru T8 at our 294Mhz point using the little PWB that could:
http://img159.imageshack.us/img159/6...nted6tl.th.jpg
A bit of improvement over the previous short term stability of T8 at 284Mhz we started with, eh? :p:

Further updates as things unfold... BH5 testing now in progress ;)

Peace :toast:

:lsfight: * Round 4 - UTT BH5 and the Java Junky * :lsfight:

Now for a little UUT BH look-see :D Decided to use an average pair of UUT BH with a similiar Winnie... and of course lots of :coffee:

First we have some baseline testing to lay out... let's see how it went:

A little low voltage (3.2 in BIOS, 3.28 measured) T5 baseline:
http://img393.imageshack.us/img393/1...242m5tz.th.png
Not bad, but then Vtt is pretty stable during T5 ;)

Let's try and crank up T8 at the same voltage/frequency:
http://img393.imageshack.us/img393/2...242m7vd.th.png
Arrrrgggghhhh! Lock-up bigtime :shakes:

After half-a-dozen reboots to lower the frequency, we arrive at a lowly 200Mhz:
http://img393.imageshack.us/img393/4...200m5sk.th.png
and while she doesn't lock-up, we STILL have errors :wth:

Ok, let's try it with a bit more voltage, this time 3.5V in the BIOS, 3.57 measured, first at 249Mhz:
http://img144.imageshack.us/img144/1...249m4dv.th.png
Nope.... some errors.

A clean run down a notch to 248Mhz:
http://img144.imageshack.us/img144/3...248m6ju.th.png

After the previous debacle with T8, decided to start low... here's 211Mhz:
http://img144.imageshack.us/img144/9...211m4zx.th.jpg
Well, at least it didn't lock-up :shrug:

Drop it down a notch to 210Mhz:
http://img144.imageshack.us/img144/7...210m7ir.th.png
Ahhhhh... an actual error free run of T8 :rolleyes:

Let's take a peek at what Vtt looked like with the MB in stock trim trying to run T8:
http://img144.imageshack.us/img144/1...20mv0is.th.png
Worse than the TCCD runs... 120mV drops :sick:

The obligitory annotated :scope: capture:
http://img144.imageshack.us/img144/7...5327vt87dk.png

Darn stock configuration is kicking the UTT BH5 in the nads :hm:

Alright, it's time to beat this MB into shape :cord: Hooked the little board that could back up :hehe:

After enduring sooooo many runs to determine the baseline conditions, I said the heck with it, and also cranked the multiplier up a notch to 9X :sonic: so I could get the rest of this round of testing finished today... so sue me :p: Everything else however, as before, remains exactly the same in the BIOS.

Now let's take another peek at Vtt while running T8 using the UTT BH and see how it looks with augmentation:
http://img144.imageshack.us/img144/3...5357vt80yb.png
Once again, we have success with the little board that does :clap: Now that Vtt looks good, maybe we can get somewhere...

A quick run at the 3.2V BIOS setting:
http://img233.imageshack.us/img233/2...nted5sg.th.png
Much better, 242Mhz and passing all tests (compared to 200Mhz and still failing) :D

Now some runs with the 3.5V BIOS setting...

A clean T5 run at 252Mhz :
http://img233.imageshack.us/img233/9...nted5rr.th.png
Pretty good, gained 4Mhz on T5 :)

But... will T8 still hold up:
http://img144.imageshack.us/img144/1...nted6dp.th.jpg
Yup, a nice little run at 255Mhz, 45Mhz above what we could do in stock trim :D

And here's an almost 2 hour ramble through all the tests at 252Mhz:
http://img292.imageshack.us/img292/5...nted4fc.th.jpg
That's 4Mhz above what T5 ran at before and 42Mhz above what T8 would do :woot:

Being the curious sort, decided to check what the peak augmentation current during a T8 run:
http://img233.imageshack.us/img233/5...rent8uz.th.jpg
Having to pump almost 3A to get rid of those 120mV droops :shock: No wonder we have some small fluctuations of Vmem at the input to the little board using those foot long wires and little SMD clips to bring power over :lol: Even with all that though, it still holds Vtt to only an 7mV fluctuation.


Had some time left, so decided to go ahead and see what happens when we ran OCCT using the handy dandy little board on Vtt...

First, a little refresher picture from the original thread on this problem of what Vtt looks like during OCCT with the MB in stock trim:
http://img146.imageshack.us/img146/4841/occtdip21ek.png
Nasty, over 200mV dips in Vtt with the MB stock. Needless to say, OCCT use to fail like T8 did ;)

And what does Vtt look like running OCCT with the addition of our little PWB?
http://img202.imageshack.us/img202/8...gmented5jt.png
Vast majority of the time, even tho working like mad, it keeps the fluctuations to under 8mV :D

There were some occasional times it had to open the throttle up full bore tho like this:
http://img202.imageshack.us/img202/2...gmented5hk.png
Resulting in a whopping 23mV peak :rolleyes: Still waaaaaaaay better than the >200mV nastiness from before... and not bad at all considering the test lash-up ;)

But.... is it enough to get us through OCCT without problem? Hmmmmmm......
http://img292.imageshack.us/img292/8...08103pb.th.png
Ok, ok... so we lost the grand total of 3/4 of a Mhz from our Memtest speeds, so fire me :p:

/me hears a voice in the background

What's that Michal... did you say Spi32M for you???? :rotf: Here ya go bud, just remember, it is NOT a max speed run, just a comparison point to the stock MB ;)
http://img292.imageshack.us/img292/8...08109mc.th.png
Even took a pic of it just for grins and giggles :lol:
http://img292.imageshack.us/img292/9...nted6dp.th.jpg

And while I was putting this post together, I ran one last little test to verify stability ;)
http://img232.imageshack.us/img232/5...934p3px.th.jpg
Just a wee 934% run of the Windows version of Memtest :p:

Sooooooooooooo... in summary:
Stock DFI nF4 MB and UUT BH, with 3.5V in the BIOS - a "staggering" 210Mhz max stable in MemTest8... :slap:
And with the little board that could - a nice 251.3Mhz stable in anything you throw at it :bananal:

Oh... and the 3.8V or higher pics... well, if you go back up to the first post you can see the board was initially tested with a 4V input ;) I dislike going over the mid 3s for Vmem on air cooling...and all this testing has been done on an air cooled low-grade Winnie at under 1.6V Vcore. Just think... with a different CPU, some proper cooling, and higher voltage... :hehe:

Have to give a big hand to ImageShack for Free Image Hosting


Peace :toast:

ps. Further updates will end up down below :)

PM me if you want me to send you a 9173A to use for testing augmentation.

That is very annoying that the current limit is tripped. It's set inside the chip, we can't... O'Clock it :slapass:

So your circuit makes up for a clear deficiency in the 9173. I was going on the assumption ( always bad ) that the 9173 internally had a good feedback path for maintaining Vtt output to match the desired Vref input. You've proven the internal 9173 circuit doesn't have good enough noise rejection to deal with the application... AND... that it cant meet the current requirements of the Vtt termination.

Hopefully adding current augmentation to your circuit can solve the last issue. Good catch EMC.

Im so used to OVER designing things, it amazes me when a high volume device...like this motherboard (100's of K's of units shipped)... doesn't have plenty of "margin" to avoid disasterous support situations.

RikTek.... you're FIRED!!!

I prpose for the final permanent design, that you use a connection to the JP17 as a way of getting at the 3.3V and Vmem sources... that way the user just needs to pop off one of the JP17 jumpers and plug on a cable from your board.

Try to minimize the soldering required.... but soldering WILL be required to tie into these high current signals when you add Vtt supply augmentation.

I'd suggest using a second 9173A only because it will most easily mesh with the onboard one. Otherwise I'd say totally supplant the onboard 9173A and make a device with Vtt regulation at higher current. Just have the user desolder-bend up 9173 pins so its not driving the Vtt anymore.

Kinda sux that we'll all have to mod our boards so extensibly just to get it to work right.

Quote:

---

Originally Posted by EMC2

Doesn't look like much, does it? Well... let's just say looks can be deceiving :hrhr:

---

:p:

Peace :toast:

Post 2 now populated :toast:

Looking at that thread makes my head hurt, seriously! But I finished it, with help of a couple of Aspirine tablets :D
I think we should create something like "electronics engineer corner" here or something of the likes, just for you Daniel :up: All that work and testing you put into it is just tremendous!

So, it looks like your little "vtt circuit" gadget fixes some of the problems and allows your TCCD to run better cleaner! Well done :toast: Looks like all that java wasn't wasted :p:

If this is a universal (Meaning it can work on more than just the DFI NF4 mobo) VTT circut add on that can be soldered and not jumpered in I will pay alot for it.

Nice work Dan,

So clearly the 9173A's regulation efficiency and noise rejection sharply falls off with output current. And in DFI's design it is severely overstressed.

As I thought, we are all leaving 10-20Mhz memory-FSB on the table with this Vtt flaw. I also believe that we are all required to use higher Vcore to achieve Prime/SuperPI stability due in part to this Vtt issue. I think you'll find that you can drop Vcore a bit on your system now, and maintain stability.

In part people getting 260-270Mhz out of BH5 or VX had boath good ram AND thier specific boards had a less tilted Vtt output. While most of use were in the 250-255 range with ram that "should" be doing better.

Dan is going through the kind of analysis that the manufacture is SUPPOSED to go through before making the product. Design review, stress testing, etc. It's pennywise and pound foolish to forego doing extensive testing at the pre-production phase, where this limitation would have been caught.

Let this be a lesson to DFI and all others.... there is no escaping your homework. Electrons are electrons :fact:

Hurry up and sign me up for one Dan !!!!!!

This looks really good and professional. Are there going to be any test with some BH-5 @ around 3.8 or so volts to make sure of this?

*Michal*
/me sends a nice Tylenol Laced "Super-Pi" your way :lol:
Which thread, this one or the original (guess is original)...

No, no.... don't put me in a little box in the corner bro, I wouldn't fit :p: Besides, I need room to roam and have just a wee bit broader range than that ;)

Thanks, but truth be told only put in about 10 hours total on the brain storming, circuit design, and PWB once I decided to do it... that part was easy. It's the fabrication, documentation, and more importantly proper testing that eats up so much time! (and the darn waiting on supplies, lol)

Oh... and I never waste good java :p:


*Craig*
In answer to the Q, if it has Vtt on it.... AMD/Intel, DDR, DDR2, big, little... doesn't matter, it'll do what it is intended to do :D Does have a number of little things I haven't gotten into as well :hehe: Along with two special "features" tailored to the DFI nF4 (but which are not limiting factors for use anywhere else). At a later point I'll lay it all out ;)

Thanks Jim,
Just FYI... it's Daniel bro ;) While the RT9173A does not have the best regulation possible, the part itself really doesn't have a noise rejection problem per say... again, mainly an implementation/usage issue here (tho it would still fluctuate 30mV or so even if it wasn't overloaded in this MB design... and maybe others *cough*nF3 maybe*cough*).

As far as other things besides max OC (will vary depending on memory and MB), I strongly suspect that a lot of the sometimes "flakey stability" people see is related... and yes, there exists a possibility that the same OC with a little less Vcore may be possible (from my experience with AMD A64s, Vcore does help a bit with the mem controller speed). There's also a feature of the A64 that isn't used often... that can be taken advantage of with this design ;)


*moddolicous*
Testing is far from complete... and yes, in fact some "new" BH-5 is next up in the testing Q.

Peace :toast:

this is an awesome thread even if i might possibly only understand a fraction of it...

But when i did my personal tests a while ago with the cheapest DMM possible and i could literally watch my OCCT crash on the DMM i *knew* we have some serious issues here and thats why i am following the developments like a hawk.

For some users its maybe 20mhz more....for others it MIGHT be the reason for general instability/flakyness.

MY QUESTION now, what do we "average" people need (on parts etc.) and what exactly has to be modified/sodlered on the board ?

I am not even interested in the BH-5 etc. tests anymore because what i saw above is PLENTY to make me want to have that little board NOW :)

Nice job You really know your stuff

Flexy -
If there's something that you don't understand, just ask ;)

One thing that might help some people is how to "read" the o'scope pics. Just remember, you are looking at voltage with respect to time. Most of those pics encompass a 20 second period in time, showing what the various signals are doing over that time frame. Time is from left (earliest) to right (latest) and voltage is up and down. You can see the scales along the bottom of the captures for the various signals and the time scale. One "division" is the area between a pair of the light yellow dotted grid lines on the scope screen. There are also "tick marks" that are spaced at 1/5th a division interval. It's just like looking at a graph in Excel or a graph in math with X-Y axis ;) Time is on the X-axis (left to right), Voltage on the Y-axis (up and down). Then the various signals have an "offset" applied to them to position them on the screen. Note that each signal has it's own offset and scale. (maybe I'll annotate a pic about this for clarity)

Regarding the other Q... when all testing is done, I'll get into that part. Regarding MB changes... pretty much like any other volt-mod... soldering some wires to a handful of spots... at most lifting 1 leg on the RT9173A or its equivalent on other MBs (may/may not be required, depending).

General comment... this isn't and never was about $s... mainly it was the fun of the challenge ;) Other constraints going into it was

• a) keep the cost down
• b) flexibility
• c) KISS (keep-it-simple-)
• d) reliability
• e) the most accurate, stable Vtt possible given the other constraints
• f) ease of use

So far it's hitting bullseyes :D Oh... and I'm included in that "we" of yours :p:

Regarding the "NOW" part... I could go ahead and say, "Hey, it works, let's go"... BUT I am a very strong advocate of detailed, documented, controlled, thorough, OVERtesting before anything leaves my big paws. :fact: I won't budge from that principle.

--- All ---
Thanks :toast:

ive researched a few ic's for this problem if you want i can send you details on the chip and you can overlook it! jlmk

Thanks... but looked at a good dozen possibilities... and the ones I'm using fit my technical requirements best... and are working very well ;)

That said, if you want my view for something you are doing, I'm always willing to lend a paw with an open mind :toast:

Sry Emc2, Even with my super dooper ultra vision I can't id all those little components you used from those pics ;)

I wonder why not :wasntme:

He blurred them out. I was going to build one on my own, but noticed that.

Awesome thread with great results, proves my point there is more to overclocking ram that more voltage.

Great work!

Thanks bro :D And your OC'ing comment holds true for many things BT ;)

BTW... look up, another round of results have been added...

Peace :toast:

Interesting results from part #4, Daniel :up:

All we need now is a couple more of those "little boards that could" for some further "beta-testing" :D It still makes me wonder If only your board (or some low overall percentage) was one that needed to be "enhanced" or all boards will benefit in same way your did?

Excellent detective work, once again! Thanks for 32M, too :toast:

Awesome work bunky.

Now Im dying to find out what my system will do FIXED. If Im at 250Mhz 512Mx4 with a stock board, will this give me 10+Mhz more after it's cleaned up!!!!

The BH5 behavior is eeirily familiar. Makes me wonder of the NF2 design has a similar flaw in it, the old BH5 "should" be doing 250Mhz and only does 200Mhz stable..... brings back bad memories.

This may be the root of why my OCZ ValueVX BH5 does so less MAX fsb than the OCZ 3200Gold BH5...all UTT BH5, but maybe the ValueVX sticks are more sensitive to the Vtt droop. It's Test8 and Windows that errors out at lower FSB. Further, the OLD BH5, like my HyperX 3000 might perk right up, as they act JUST LIKE your BH5 on stock board.

Looking forward to your getting some UTT-CH/VX sticks for testing....just to see what level Vtt error it generates, and the same for Ballistix sticks.

Sign me up for one LBTC from the first batch.

First, a couple of more pics of note....

While I went about life, left the board running Windows Memtest and ended up with this pic:
http://img219.imageshack.us/img219/9...test0wl.th.jpg
Over 10,000%... pretty darn stable, eh? :cool:

And here is a pic of what the augmented Vtt looks like during a Spi32M run:
http://img386.imageshack.us/img386/4...i32mvtt1pe.png
Based on the augmentation signal, it too is affected by Vtt.

--- Michal ---
Regarding wondering "if it's only my MB"... or if "all boards would benefit as mine did"...

No, it definitely isn't just my MB. I doubt mine is the worst (in fact suspect some get returned as a result of this as "unstable/bad" MBs). In addtion, I'm quite sure there are others that don't have the same magnitude of problem. Just like any other electronics, be it memory or CPUs or other components, there will be a distribution of values.

But... I would for starters refer you to the original thread where I raised this issue.

Here are some pics Oskar took (#120). One is with the MB in standard trim, the other with the Vctrl input to the RT9173A tied to the +3.3V rail. You can note that while he has them "compressed" using a 100mV/division scale, they too fluctuate somewhere in the 60mV to 90mV range.

Then take a look at Oskar's comments in this post of his (#124). Note in particular his last comment... 4 boards looked at, all act close to the pics he posted... and his next to last comment, in item 3 that his measurement of 94mV is not so good...

How much it would help would vary from system to system... and condition to condition. Running 1 DIMM or even likely 2x256 it would make little difference for example. For a given MB, the more memory used, the faster the speeds, and the tighter the timings, the more it could help. In some cases it might show up as "wierd" intermittent instabilities as have been seen. In others it might be max OC limitations... or the need for slacker timings (example: I can actually run the memory higher with the MB in stock form, if I slack the timings waaay off... do things like 2T with only 2x512MB, no interleave, etc., etc.) And it is not just with the DFI nF4. There are other MBs out there with this issue, one example might be as uwackme raised above...

You can get somewhat of an idea by various methods without a :scope:. One is to take a good DMM to Vtt while you run MemTest 8 (or OCCT when it fails). Another is to compare your max stable Memtest T5 and T8 runs (30mins to 1 hour runs, not 2 pass runs)... or your max stable Memtest runs to your max stable windows benchie frequencies.

I know there are those that will/do disagree, but Memtest is actually a very good indicator/tool when properly used. You'll notice for example, I lost less than 1Mhz between stable Memtest and stable anything in windows. Historically over the years, I've seen from none to only a couple of Mhz difference between the two (memtest and windows) on a properly setup and tuned system (this includes things like timings, PSU, MB tweaks, etc.). When there has been a greater difference, it has always indicated a problem of some kind (not necessarily MB). The same can be said for T5 vs T8 in my experience regarding indicators... T8 usually can be run at a higher memory frequency than T8. Which test fails and how can be a very useful tool ;)

Would everyone gain the large amounts I did... NO. Would some gain more... yes. Fact is, I would have shown greater gains with a different CPU/cooling method ;) Others might only show a couple Mhz. On some other MB designs, you would get no gains. Or your CPU/memory might hold you back from reaching the point it would be beneficial. But... in the land of OC'ing, any gain is a gain ;)

Oh... regarding the 32M... I knew you wanted to ask :p: And, you're welcome bro :toast:


---uwackme---
Thanks :D

Most of your Qs are answered above as best can be. Regarding other MBs... quite possible, in some cases very probable. Power and termination and the like are often overlooked in "the-need-for-speed" :fact: As Bigtoe alluded to, there are many aspects to OC'ing, other than cranking up the voltage. He also made an interesting comment that's semi-well known, although that is the first time I've seen it publicly quantified...

Quote:

---

Originally Posted by Bigtoe

FX55 have the best memory controllers so overclcoking ram is a little bit easier. our tests back at OCZ have shown an FX can be anything upto 25fsb better with ram over an el cheapo venice or winchester.

---

I have little doubt that with, for example, an 0442 FX the mid-range UTT BH I have would push quite a bit higher.

The UTT-CH testing might happen ;) As far as the Ballistix sticks - anybody care to donate some for testing? :lol: I will say this tho - regardless of memory type, the affect is there once you go to two or more double ranked DIMMs.

LBTC previously noted ;)

Peace :toast:

a compare to hipros booster would be very interesting

my pc overclock stable super pi memtest prime95 at 255x10 but more than that give erros in all programs,test 8 memtest sometimes give error from 258 to 265mhz,test5 only give error more than 265,i think this could be cause by this problem,so making the mod is the only way to find iff overclock more and more stable.
my cpu can do 2800mhz(xp-120) at 1.475v but give errors too(255x11)

This explains the behavior I am seeing with my Redlines. Passes test 5 no problem but test 8 will fail at random intervals during a 100 pass run. Looking forward to the fix for this.

--- blander ---
Short answer - two completely different critters based on what I know of George's Maximizer... their only commonality being they help you to OC your memory higher ;) The only thing the LBTC does is either control, augment, or replace the actual Vtt voltage on the MB and must have its connections soldered in (test clips are fine for testing or benching if careful, but too easily dislodged for "real use" IMO).

--- Portugal ---
To tell for sure, drop your CPU multi down to 8... but if you can run T5 for a long time (hour) at 265 without any errrors and you get errors in T8 above 258 and errors in windows and things like Spi32M, OCCT, etc. with a lower CPU multi, then the answer most likely is yes (take it as a given that you have your timings properly sorted out). Prime95 I can't speak too, as I don't use it usually.

--- Q4S ---
What happens during Spi32M or OCCT? T5 speeds or T8 speeds?

Peace :toast:

BTW... 3.6V BIOS gave 255M Spi32M tonight

so,here is the results from tests at diferent multipliers:

264mhzx8 no errors in test 8(tested 1hour)
264mhzx9 no errors in test 8(tested 1hour)
264mhzx10 9 errors in test 8(less than 30minutes) here is the pic:
http://img300.imageshack.us/img300/6577/test89ex.th.jpg

i see the errors is in diferent adresses and mb in all loops,so what thinking about that?

I just finished testing and concluded that the VTT fluctuations is limiting the memory OC.
For the testing I Used SuperPi32, Sp2004 & Memtest.
I was hitting a wall at 265 with any voltage. Errors in memtest 8 and fails SuperPi 32M
After I did the VTT mod and raised the voltage .02 it now runs clean at 265 :woot: and passes all tests without errors. I am now testing at 267.

I know the vmod is just a band-aid so I am looking forward to the fix.

Ive been having issues with my BH5.. Passes test 5 fine at 260-265 but errors out test 8, occt and 32m superpi in a bad way.. It doesnt seem to affect my TCC5 much though.. This same BH5 runs great and doesnt show this when running in my 250gb mobo..

Looking foward to your fix EMC2.. :)

I don't want something that I plug into anything. I want to be able to solder it to components on the mobo. I hate plugs. I solder wires to everything, the back of my power headers on my mobo and my videocard power header, sometimes directly to capacitors or chips, depending on the application.

I support the solder only installation of it.

Just a few random items to add.

The 9173 datasheet includes some weasel-words about how the customer is "encouraged" to duplex this part to reduce power consumption in dual-channel DDR SDRAM applications (like the NF4). As a former applications manager for a semiconductor company, I can tell you that "encouraged" nearly always means "we're out of margin - do this if you want it to work". The NF4 board design probably would have included two of these puppies if RT's advice had been trusted.

The SOP-8 package version of the 9173 has the same Rds(on) as the TO-263 package, which means RT is using the same die size for the two parts (but current limiting the SOP-8 for thermal reasons). In a package the size of the TO-263 they could easily fit a MOSFET with half the Rds(on) of the SOP-8. Looks like they'd rather sell you two parts than use a bigger die. Cheesy. :rolleyes:

ChiefTech/Fortune is working on a pin-compatible replacement for the RT9173A, but it's not sampling yet (FS8863B). Might be worth following.

sluggo

So do we think just piggybacking a second 9173A would be helpful?

you know i seen that on the datasheet, so this makes me wonder will the fluctuations still occur out of dual channel? personally i dont see much differance between the two for power consumption but it might be worth a look.

Quote:

---

Originally Posted by tennvols_69

you know i seen that on the datasheet, so this makes me wonder will the fluctuations still occur out of dual channel? personally i dont see much differance between the two for power consumption but it might be worth a look.

---

that would explain why memory overclocks are higher/more stable in single channel mode then dual channel right there haha.

also ive read many post that say something like i got two sticks of ram that do 250 2-2-2-5 each but together my system run stable even at stock settigns of 200mhz and im thinking this might have something to do with it on the more unlucky boards
________
Honda jazz specifications

btw i forgot to mention emc2 you are my freaking hero man i think i will try to improve my knowlage of electronics. hey you need to be pretty darn good at math to become an electrical engineer? anyways im real gratefull of your work just ask if there is something that i can ever help you with
________
Zx14 vs hayabusa

Quote:

---

Originally Posted by tennvols_69

you know i seen that on the datasheet, so this makes me wonder will the fluctuations still occur out of dual channel? personally i dont see much differance between the two for power consumption but it might be worth a look.

---

It's a worthwhile question. I don't know the answer, but it seems to me that since dual-channel operation will increase the duty cycle of this part (which, aparrently, was designed only for single-channel applications), the part willl be running hotter than it would be in single-channel. If it's current limit is appropriate for single-channel use, then it should also be fine for dual-channel, as only one channel is active at a time. If, however, the part was only characterized for single-channel and it's running hotter in dual-channel, then the thermal and current limiting data on the datasheet may not be applicable for the NF4 design.

Quote:

---

Originally Posted by uwackme

So do we think just piggybacking a second 9173A would be helpful?

---

I'm not sure I'd be willing to try it with my own board. It looks like it might work, but RT recommends having the two parts drive alternate lines rather than using them in parallel. I'm guessing they recommend doing it that way for a reason.

sluggo

subscribe :clap:

More random info. The early version of the datasheet for the 9173 called for a 100uF cap on the 2.5V line, while later versions call for 470uF. The later spec also adds a 1000uF output cap to the original 10uF (!) recommendation.

I looked on the board to see what DFI did about RichTek's changes to the reference design. They went 'em one better - DFI added two 470's and two 1000's. I guess they had some concerns even back then.

sluggo

Wow huge hand for EMC2!!! :clap: :clap: :clap: :clap: Thanks much bro! Why arent you an XIP yet?!?!?

Now, what do I have to do to build one of these gadgets? Maybe a scematic? Also, were can we get those PCBs you have or can we get them from you?

THANK YOU! WOOOOOT FOR EMC2!!!! :woot: :woot:

(I wount tell you I love you as last time you said "But if you ever try to kiss me again, I'm gonna use a LEAD PIPE on your head!") :rofl:

ok i'm gunna play devils advocate.

what rev of dfi board is that ?
how many different DFI boards did you test ?

no board is perfect they all got nasty drops in vdimm and vcore some worse than others.

but you said you only got like 210mhz on stock DFI board with UTT, that man is just the worst mobo i have ever seen. no one i have seen was limited by that much on a dfi nf4..
now that being said....

omfg!!!! what a brain bender of a thread!!! thats some seriously nice work youve done!

ok back to play devil's advocate.

not all the boards are like that i really believe you got a massively crap dfi baord. that hated that utt and took some pretty massive vdroops.

but then again WOW!! i just cant say that enough WOW...

My board performs similarly. I have a much lower end scope and I can watch where it drops as it switches from off to on. It's not just his board.

WOW, just wow.. exellent engineering work there emc2 :clap:

I will do some testing, the following week and see if this could explain my problems with S-Pi, it wont run with high mem speed (bh-utt) whereas it is totally memtest5 stable (i have never bothered with test8)..

Will it be possible to buy the solution in europe/norway when your testings are complete?

Id also like to ask one (probably) stupid question :p: what does java mean?

Java = Coffee bean

--- Portugal --- You have a several week old PM ;)

--- Craig --- If I do, it will be however one wants to do it... flexibility is always a given ;)

--- Uwackme --- As I said before, it might work... question would be the differences in the Vos between the chips.

--- tennvols --- Go look in the original thread I referenced... a single 1x512MB doesn't create the problem. The loading is half ;)

--- brandinb --- Yes, lots of math is a requisite... geometry, trig, calculus, differential equations, partial differential equations, etc. are required for any engineering degree. But, you can learn a lot about electronics without it... just limits you in some areas without it.

--- looks at Stealth and brandinb and taps keyboard with pipe --- :p:

--- sluggo --- No, in DC, both "channels" are active at a time... it's a 128-bit memory bus configuration rather than a 64-bit configuration ;) Regarding the output capacitance... it helps with *short term* fluctuations, supplies current until the regulator catches up... but does nothing for long term fluctuations like the nF4 has.

--- Lestat --- the limitation depends on CPU/memory/timings and board. Regarding if it's only one board... go back and read the references that I gave to Michal in post #24 of this thread and my own comments playing advocate ;)

--- un0kai --- if it goes to sale, yes... BTW, there is no such thing as a stupid Q when it is an honest quest for knowledge ;) In fact, such questions are smart ones :)

Oh... and regarding Java:

Quote:

---

"It is by will alone I set my mind in motion. It is by the juice of Java that thoughts acquire speed, the lips acquire stains, the stains become a warning. It is by will alone I set my mind in motion."

---

Fantastic stuff EMC2, takes me back to my uni days! Wish i had a scope here to do some playing myself. Keep up the great work!

and wath about trying to ad output mosfets to get more current?

do anyone know if dfi have fixed this vtt problem on the new revison (dr-extra) board? My guess is it is still there?

and btw. i love your avatar craig588

and emc2: regarding the quest for knowledge, I could not have said it better myself...

First up a massive hand to EMC2, brilliant stuff, well done! :toast:

This is the kind of overclocking stuff I like to see, people putting to use electronic engineering knowledge to improve overclockablility (not like the "ADD MORE CAPS!" approach :stick: ).

Second, are you planning to make and sell this device or just release the a circuit digram or both? Sign me up for a copy of the diagram if you release it. :cool:

Edit: I had a little thought, I wouldnt mind paying $5-10 for the diagram and a little guide, just to justify your time and effort spent on it. Doesnt sound like much, but given this forums prominence in the world of overclocking, you could be sell a few thousand copies of it easily.

OK, i am from Germany, i really don't understand what EMC2 has done there Could someone explain this a little bit for me? What do I have to do? :confused: And what gets better than before?

Any news?

Just looked at this thread and it is VERY interesting. Wish I can buy one.. :(

Quote:

---

Originally Posted by Rolf

OK, i am from Germany, i really don't understand what EMC2 has done there Could someone explain this a little bit for me? What do I have to do? :confused: And what gets better than before?

---

The memory busses on this board use active termination. The dimms must be able to pull up AND down in order to transfer data. The termination voltage is called Vtt. The device on the motherboard that supplies Vtt can sometimes lose regulation and Vtt then droops as much as 200mV. When this happens, data can be corrupted.

EMC is working on a circuit that will force Vtt to track Vdimm more accurately and thus allow for more reliable operation and higher overclocks.

What you have to do is wait for EMC to say "it's done". What happens after that is unknown at this time.

Quote:

---

Originally Posted by sluggo

The memory busses on this board use active termination. The dimms must be able to pull up AND down in order to transfer data. The termination voltage is called Vtt. The device on the motherboard that supplies Vtt can sometimes lose regulation and Vtt then droops as much as 200mV. When this happens, data can be corrupted.

EMC is working on a circuit that will force Vtt to track Vdimm more accurately and thus allow for more reliable operation and higher overclocks.

What you have to do is wait for EMC to say "it's done". What happens after that is unknown at this time.

---

Thanks, that sounds very very interesting! :fact:

Wow EMC2, nice job you did there, how is the design of the circuit going? :)

OK, that's what I found today by chance:

http://www.micrel.com/product-info/SSTL-Term.shtml

This sounds very interesting. :) 5A would be really nice when running 4 RAM modules.

Any chance EMC could give the once over to the new DFI ATI Crossfire motherboard? just to see if Vtt is not an issue. Dont even want to consider buying it unless its clean of issues like this.

Quote:

---

Originally Posted by celemine1Gig

OK, that's what I found today by chance:

http://www.micrel.com/product-info/SSTL-Term.shtml

This sounds very interesting. :) 5A would be really nice when running 4 RAM modules.

---

This looks very interesting!
Any updates EMC2?

please emc2 speak to the masses!!!

curious to hear your update
________
Problems from zoloft

i couldnt wait any longer for the board that could.
found this baby:
http://pdfserv.maxim-ic.com/en/ds/MAX8554EVKIT.pdf

this board can be setup to run vtt from what i understand provide upwards of 8A and 550KHz switching freq. they want 60$ shipped in the usa.

i can begin to figure why they have such a wimpy part for vtt on this board.
3A is minimum for ONE DDR channel. it didnt take me long to find this out reading white papers. shame on DFI.

Where can I buy it?

just as an update to my previous post

i have 2x512 BH5 sticks

could not run stable at any speed before
even one stick seemed iffy

after adding parts to adjust the output volts (7K resistor and 10K pot)
running the board off the 12v line
using 200Khz switching freq
(rt part on nf4 apears to run a 1khz freq)
http://www.maxim-ic.com/quick_view2.cfm/qv_pk/4247

now @245 tight timings its rock solid.

I dont have any scope that would be useful here, my dmm shows no more than 20mv voltage swings. the output coil gets very hot, i have a feeling this is due to the length of the hookup wire. (about 6 inches).

This seems very interessting and all, but I'm not good enough at electronics to understand this. Is there any way I could use some of these stuff for my neo2 platinum of msi, my ram just isn't running at what it's supposed to be able of.
If someone could write me a guide... (I know it's lame not figuring it out myself, but I'm really a noob in electronics :()

Gosh, I can't wait for the solution here.

EMC2's last post on these forums was almost 1 year and two months ago. I'm tired of waiting for this mod however. Does anyone know anything about how it worked? To my understanding the base of the circuit was a difference or differential amplifier.

From one of EMC2's posts in another thread. (I was reading this wrong earlier, the Vref, Vref2 and Vtt aren't the labelling of the DIFFAMP but the labelling of the connections from the RT9173A).
http://xtremesystems.org/forums/atta...5&d=1121305069

Instead of having the diffamp connected to pin 4 and pin 5 for the inputs, connecting it to pin 1 and pin 5 would give us the desired results. Otherwise you need a different circuit where Vout = 2 x vref - vtt (and Vref = 1/2 Vmem, so using Vmem seems a bit easier).

However, for this circuit to provide 2.29v to pin 4 (EMC2's device's result in one of the graphs on the first page), Vtt would have had to drop to 0.51v which is unlikely. The graphs suggest that a difference in Vtt ofaround 90mV resulted in a 900mV shift in Vref input (a gain of 10). Meaning that the equation required is Vout = 10(Vref - Vtt) + Vref. (So for every mV that Vtt is out of range, Vref will be adjusted 10mV, hopefully resulting in Vtt tracking Vref more closely).

Edit: Supposedly a 'passive averager' circuit could be used. The output from the diffamp and Vmem can be averaged using two equal resistors. The resulting equation would be: Vout = (20(Vref - Vtt) + Vmem)/2, which happens to equal the target equation (Since Vmem = 2Vref). I can't find any evidence that EMC2's circuit works in this way, but the end result should be the same. This leaves one semiconductor device for the 'augmentation signal'.

Edit2: By the looks of things, most diffamps / opamps cannot provide more than a fraction of an amp. The 'augmentation signal' was pumping out almost 3A peak, so I'm currently looking at different methods of power regulation. It should be possible to use the output from the 'passive averager' as an input for the power regulation to track. If the power regulation method can output 3A is should average out with the Vtt output from the RT9173A, but another passive averager might be needed.