512mb 7900GT vmod development thread

[OneProdigy]

That would mean to desolder the resistor and add a higher rated one instead, right?

EDIT: I think i must have mixed up that 1/2 vmem thing with vtt, so nevermind.

[alpha0ne]

ok guys the solution to this little puzzle is to buy a sparkle calibre 7900gt 512mb.
the vreg design is the same as the 256mb 7900gt's and they should clock similarly.
i'm vmodding one within 48hours and will post results.
the core vmod is the same digital vmod as the 256mb cards and the 1.4ns memory seems to clock better than the legend/gainward style pcb's.

there will also be the same issue of risking burning the memory vreg if pushing the memory too hard.

I have been very seriously thinking of buying the Calibre card for the last week but unfortunately they dont use the revised PCB (even though it is VERY minor), I'll make up my mind in the next day, otherwise it will be another MSI except VIVO and revised PCB

[sluggo]

Would make sense to me that either vddq or vref (or both?) are a little off the specs...even the 512MB cards equipped with 1.2ns ram seem to hit a wall at around 1550Mhz although they should be quite capable of doing that speed.
So imho there's a general memory voltage problem with these cards...

Maybe i'm wrong...but shouldn't vddq always be half of vmem?

The memory is spec'ed for 800MHz operation (1600MHz data rate), so saying that it should operate at 1550MHz is something I can't disagree with. However, it's not obvious to me that it's a voltage problem if the card's memory does not operate that fast. There could well be other problems that limit the data transfer rate, like some of the other timings set by the card's BIOS. I'm no expert on these timings, but my guess is they are just as critical here as they are for system memory.

Vref is supposed to track Vddq, so that if Vddq is increased then Vref should also increase. The ratio is not half, though. Vref is defined as .7*Vddq, so for a 2V Vddq, Vref should be at 1.4V. Cheers

[Scarlet Infidel]

So, glancing through the datsheet im going to summarise a few things, mainly for my sake.

....................Should Be .......................Measured

tRC.....................35........................ .........34
tRFC...................45......................... ........43
tRAS...................25......................... ........24
tRP.....................10........................ .........10
tRCDRD................12.......................... .......12
tRCDWR................8........................... ........8
tRRD....................8......................... ..........8

VDD.................1.9-2.1v......................... ~2.1v
VDDQ...............1.9-2.1v..........................not sure
VREF.............1.449-1.491v..................maybe 1.24v (for 2.1v VDD)

Sorry about the appalling formatting. It seems most things are within specifications except the timings are a little tight (shouldnt matter as they have set VDD at its maximum and the given timings are meant for 1600mhz anyway). This VREF definately needs more investigation.

The data sheet says VDDQ must be less than or equal to VDD (also should be between 1.9v and 2.1v as with the VDD), does it necessarily run at the same voltage?

Also VREF is 0.69-0.71 times VDDQ. Is seems that whilst the VDD is fine at 2.1v and the VREF is actually at 0.7 times the VDDQ as its meant to be just that the VDDQ is far below the VDD.

If the VREF is 1.24v then this divided by 0.71 (potentially) gives a VDDQ of 1.746v, is that what was measured at 1.74v? has anyone verified the assumption that the 1.74v was not DC as suspected?


I want to see someone raise their 1.24v in relation to their VDD. Then id be interested to see if 1, the memory clocks any better and 2 if the 1.74 or any other voltage changes.

OneProdigy if when you have your card you could measure the three voltages i have mentioned so i can see what your exact card reads. Thanks.

Edit: did anyone ever find more details about the NCP5424G. Ive looked over the datasheet for the NCP5424-D. Im not sure how it differs and what the suffixes actually mean.

[sluggo]

Hey Scarlet - good data. Vddq is the voltage used to power the output drivers on the chip, while Vdd is used for the internal logic. The restriction that Vddq must be less than or equal to Vdd is there to prevent a condition called "latch-up", which can occur when a transistor is controlled by a voltage great than it's rail voltage.

MASSIVE EDIT: I keep forgetting we're talking about the BJ11/12 512MB part.

A Vdd or Vddq of 1.74V is not in the ball park for the 512MB part. The minimum voltage allowed (according to the data sheet) is 1.9V for both Vdd and Vddq. For the lower speed parts (BC14/16/20), 1.7V is the minimum, but that's not what this board uses, as far as I know. Please check the board for the part number suffix and correct me if I'm wrong.

The suffixes for the NCP5424 are just packaging fields. "D" is the standard part, "DG" is lead-free packaging,

[czuk]

So according to what sluggo said, Vddq shouldn't be the solution for the memory issue
I just don't get it, why a 1.2ns memory chip doesn't run at specified speeds even when voltages seems to be okay.


If it does help, one can find all kinds of datasheets at:
http://www.datasheetarchive.com/

[sluggo]

So according to what sluggo said, Vddq shouldn't be the solution for the memory issue

I'm not sure if that's what I said ... at least, I don't think I'd claim that now. Vddq supplies the interface to the chip, and it may well be that the interface is the critical path. The memory cells may be clocking along with plenty of margin, but if the interface is not controlling the bus well, then the part still won't be a speed demon.

I just don't get it, why a 1.2ns memory chip doesn't run at specified speeds even when voltages seems to be okay.

Agree, the memory is not "being all that it can be". There are so many things that can create problems, though ... layout, clock integrity, clock skew .. it's a wonder it works at all. Consider that the memory does not drive the data lines between 0 and 2V to indicate 0's and 1's. The difference between a logical "high" and "low" is just .3V on these parts - just .15V on either side of Vref (trying to drive 2V transitions would take way more time, way more power, or both). With just 150mV of signal, the sensitivity to noise is very high.

It's late and I'm starting to babble. I still haven't heard from evga on the clock_in skew situation, will post here when I do.

[Scarlet Infidel]

Well here in the UK its nice and early (in my terms) and ive finally got half an hour to play with this card. The 1.74v points on my card measure about 3.80v which pretty much confirms that they are PWM something or others. I dont have much longer but ill see what else i can find. My ram also seems to get 2.11v and wont clock higher than about 1560mhz.

[OneProdigy]

OneProdigy if when you have your card you could measure the three voltages i have mentioned so i can see what your exact card reads. Thanks.

If i'm lucky i should have it back by the end of this week.
I'll report back with some voltage readings then...

[s e t h]

results for the 512mb sparkle 7900gt: http://www.xtremesystems.org/forums/...d.php?t=101451

edit. a distributor erroneously sent me 2 more of the legend cards so i'll try a little more modding too

[s e t h]

well this second batch of legends has infineon ram
it's rated at 700mhz and on these ones is clocked at 660 default
i hate it when manufacturers change specs after the launch batch.
the reviewers get the good stuff and the first customers too.
everyone gets excited and the bloody manufacturers despec the damn product, never tell anyone, and basically wind up scamming the masses.

infineon A
HYB18H512321AF-14
FVV14003
0552

http://www.infineon.com/cgi-bin/ifx/...geTypeId=17099

ggrrrrrrrrr........

[kamakazibond]

There is issue with cooler on agp cards, well Im 99% sure the backing paper on rampads is wrong and should have been removed, this causes core to be bad contact eap at side nearest memory and had to fill with extra thick artic silver 5, if paper is removed it will be less a gap (awaiting reply by e-mail) none my other NV Silencer rev3 told me to leave paper on or the 20pipe version I had before of this card, I tihnk Gainward have made mistake.

From what I have seen of my friends Gainward BLISS GS+ 512MB, we took off his heatsink because of all the rumours about the "Yellow tape". At first b4 we took off the HSF I too thought that something had been left on acidentally... You can clearly see that there is some orange or yellow sumthing where the white thermal pad contacts the ram. But upon closer inspection.. I could easily see a cross-hatching pattern on the "tape" and instantly I was reminded of many thermal pads I've seen over the years..

My opinion is that they used a combination of think, mushy thermal padding, and a layer of denser, thinner thermal padding. When we took off the HSF I was expecting to see yellow plastic backing tape, the kind you find on any pre applied thermal adhesive.. BGA ramsinks, MEMory heatspreaders, etc. This was not the case, and frankly I'm surprised that anyone would have thought that it was the backing tape..

We took some pics.. Lemme go resize then a bit and I'll throw them up here..

[kamakazibond]

Here your go...

Pic #1
Pic #2
Pic #3

Make your own opinion I guess, but we left his on. I also can't see how that material could possibly block the GPU from contacting the copper.. as the pics show, the white mushy stuff had melted a bit, and i'm sure with some pressure that it would smush alot more... You can also see the size of the "tape", it's ultra thin..

Any I've been meaning to post some of these here for prolly a week or so, and now I have...

[alpha0ne]

well this second batch of legends has infineon ram
it's rated at 700mhz and on these ones is clocked at 660 default
i hate it when manufacturers change specs after the launch batch.
the reviewers get the good stuff and the first customers too.
everyone gets excited and the bloody manufacturers despec the damn product, never tell anyone, and basically wind up scamming the masses.

infineon A
HYB18H512321AF-14
FVV14003
0552
ggrrrrrrrrr........

To me that is no better than fraud, I'd make a point of never buying there products ever again

[czuk]

Here your go...

Pic #1
Pic #2
Pic #3

Make your own opinion I guess, but we left his on. I also can't see how that material could possibly block the GPU from contacting the copper.. as the pics show, the white mushy stuff had melted a bit, and i'm sure with some pressure that it would smush alot more... You can also see the size of the "tape", it's ultra thin..

Any I've been meaning to post some of these here for prolly a week or so, and now I have...

Many thanks for the pictures and your conclusion. I totally agree that this yellow stuff never was meant to be removed. You can actually feel, that it is kind of thermal material, as it feels quite cold imo.

What do you think about the black protective frames around gpu and bridge chip? Some users reported that the removal of these will give you a temperature advantage of roundabout 5°C.

[kamakazibond]

What do you think about the black protective frames around gpu and bridge chip? Some users reported that the removal of these will give you a temperature advantage of roundabout 5°C.

Well thats a hard one.. We left his on, but thought about reseating it.. since the placement wasnt perfect and it bulged in one area around the GPU.

I guess it depends on your own preference.. His temps are usually around the low 50's or so (from memory) so I dont think it is really an issue..

Those black "support" pads did remind me of my 9700 Pro's Metal shim I removed when Installing aftermarket HSF.. So I guess as long as your careful, you could remove them.. The argument I read was that they were holding heat in around the GPU and HSI. I'm sure thats possible.. but they also provide leveled support for the HSF not to crack the corner of the GPU or HSI, which the GPU and HSI remind me of Athlon 32bit chips, Which CAN crack.. ( I have a cracked corner on a 1800+ and IDK if I did it or not ).

So if your careful with it.. I suppose it couldn't Hurt, but I suppose you could also remove the majority of it, leaving only 4 small pieces in the corners for stability.. That would allow air to flow by and still keep the support around the GPU & HSI.

[rOOk]

What do you think about the black protective frames around gpu and bridge chip? Some users reported that the removal of these will give you a temperature advantage of roundabout 5°C.

I'm one of them. Removed the frame and gained 5 C° (61>56).
But I think that the problem could also be on the original heatsink spring based fastening system. Maybe it can't apply appropriate pressure on the (foamy rubber) frame to 'squeeze' it enough.

Later I mounted a Zalman VF900 and pressure il a LOT more....it bends slightly the card. So maybe no reason to remove that frame if a good heatsink will take place of the original one (not bad but loud).

[kamakazibond]

I think that the problem could also be on the original heatsink spring based fastening system. Maybe it can't apply appropriate pressure on the (foamy rubber) frame to 'squeeze' it enough.

That very well could be, a sloution might be to take a razor and cut away most of it, only leaving the corners. That way it would require less pressure to push it onto the GPU as well as allowing air flow under there, while still giving the intended support.

[Thomas]

I changed bios to 1.4 volt and i measure 1.4 volt

[czuk]

I changed bios to 1.4 volt and i measure 1.4 volt

Could you please upload your bios-file, so that we could figure it out how you did it?

If you need a free upload service, try this one:
http://www.updownloadserver.de/

[czuk]

For the Pencil-Modders:

Attention!

You should measure the resistor value of the pencilled resistor (or GPU voltage) after a couple of weeks of use. We (german users) figured out, that it did DECREASE and thus, GPU voltage raise significantly!

On my card it was pencilled to 1.28V and now I it is 1.33V. Another user hat pencilled his card to 1.4V and now it was 1.6V!

Just validate your values again, as otherwise that will mean the end for the voltage ICs (and GPU of course).

[MaSTeR-StyX]

For the Pencil-Modders:

Attention!

You should measure the resistor value of the pencilled resistor (or GPU voltage) after a couple of weeks of use. We (german users) figured out, that it did DECREASE and thus, GPU voltage raise significantly!

On my card it was pencilled to 1.28V and now I it is 1.33V. Another user hat pencilled his card to 1.4V and now it was 1.6V!

Just validate your values again, as otherwise that will mean the end for the voltage ICs (and GPU of course).

yes, ...
4 days ago i had 1.71 Volt, now i have 1.75 Volt !

... but, my card only do 750 / 1440 with 1.71 Volt < very bad results for this voltage i think ... -.- pff, and if the card goes RIP, ... i get a new one


P S :

have someone the v-mod for the VREF ??
... i will test it, if it could help to overclock the infineon 1.4ns memory better

[MaSTeR-StyX]

So girls,

i have decreased the resistance of the contact obove the 667Ohm




639 Ohm > 1.24 Volt < VREF (695 MHz ram)
629 Ohm > 1.21 Volt < VREF (695 MHz ram)
617 Ohm > 1.19 Volt < VREF (695 MHz ram)

>> No Effect << !

should i decrease the voltage more ?
... i thought that i have to increase the voltage and not to decrease

[MaSTeR-StyX]

Scarlet Infidel, ...
i have decreased the resistance of the contact obove the 667Ohm,
(persivore said it on post #86 )



here the changes ... but 0 MHz plus

i dont know how to increase the voltage of the VREF ...


... h e l p ?

[Scarlet Infidel]

Im afraid we are at a loss on this one, it may just be imperfect design. I havent thought about this for a while but ill have another think.

Id recommend you return everything for normal if it isnt helping you, just to prevent it causing any damage.

Anyway, question time. Did lowering the resistance both decrease the 'VREF' and increase the 'VDDQ' on its own or have you changed anything else?

Now, if i remember correctly the one you refer to as VDDQ is actually something else which isnt even DC.

Also, exactly which card have you got? A Legend? We were mainly talking about the version with 1.2ns memory (or at least i was paying more attention because thats my version) but it seems the situation with the 1.4ns is the same.