512mb 7900GT vmod development thread

s e t h · 05-08-2006, 12:48 PM

i'd like a little help vmodding my new Legend 7900gt-sonic 512mb card.
here are some pics. at this time i'd prefer smd-grabber or pencil mod options if possible.

as far as i can tell vgpu is 1.3 at the moment and the samsung 1.2ns ram is being fed 2.11. i'll confirm as i start varying the voltages.
stock clocks are 550/700 and very little overclocking is possible above these clocks.
there also seems to be a problem with coolbits and these cards as my buddy killed his within a few hours after returning coolbits to default and going out for a few hours.
when he returned the card was dead. so i reckon use atitool or something else for overclocking till we figure out what is going on with this style card

below are a couple of pics that are exactly the same as my card and i've done a little research on the main vgpu ic.

i plan at this stage to fix an smd grabber from pin7 to ground and start decreasing resistence on it from max downwards.
please somebody confirm this will work or stop me from killing the card

any other tips or tricks or information i should know i would greatly appreciate.
i have no formal training in this field so what i know is more likely to be dangerous than not

thanks for your input

green are traces. blue is ground
error correction LM35B is actually LM358 when magnified properly

PSBAJ07 could be PS8AJ07 or o7

and for reference

front

s e t h · 05-08-2006, 10:40 PM

bump. reserved for complete mods

persivore · 05-09-2006, 12:59 AM

That looks ok. I would suggest using a 50k or 100k VR to start with.

s e t h · 05-09-2006, 01:04 AM

thanks for the reply.
you sure about using 50k or 100k.
i figured since i need fine adjustment around 2.5k that i should use a 10k one.
i was just about to go ahead and try it anyway :p

persivore · 05-09-2006, 01:24 AM

Quote:

Originally Posted by flytek

thanks for the reply.
you sure about using 50k or 100k.
i figured since i need fine adjustment around 2.5k that i should use a 10k one.
i was just about to go ahead and try it anyway :p

The general rule to find the correct value VR to use is to multiply the resistance between the feedback pin and ground by 20. (giving about 48k in this case). Using a 100k VR would be safer when testing the mod, but would give less accurate adjustment for general use.

s e t h · 05-09-2006, 01:31 AM

ok well i must then thank you for saving my card

i was about to try the 10k vr before your first reply
i only have a 50kohm multi-turn here with me.
will that be safe to try?

thanks again

persivore · 05-09-2006, 01:37 AM

Quote:

Originally Posted by flytek

ok well i must then thank you for saving my card

i was about to try the 10k vr before your first reply
i only have a 50kohm multi-turn here with me.
will that be safe to try?

thanks again

50k should be fine

. A 10k VR would probably have given about 1.6 to 1.7v Vgpu, so it probably wouldn't have immediately killed the card.

s e t h · 05-09-2006, 02:14 AM

ok by stepping up slowly with atitool and without vmods 3dm06 has passed artifact free at 600/700.
i'm sure that with little extra vgpu (ocp/ovp permitting) these cores will be fine.
it's the ram that seems to be very hard to overclock.
maybe it's only getting 1.75 and not the 2.11 i measured.
at 1.2ns default rating it should hit 800 with ease but so far i've not managed to get over 740mhz.
anyway i really appreciate your help with this

s e t h · 05-09-2006, 02:51 AM

ok more info on stock volts max clocks.
clear at 600/740
625 core has heavy artifacts in canyon '06 (screen flashes, sparkles and colour errors. clear run till 'canyon' though)
750 on the ram won't even start 'proxycon'
so max core is somewhere between 600 and 625 and max ram is like 740.
time for some vmodding

here's a smoothe bench screeny for stock volts

s e t h · 05-09-2006, 06:22 AM

ok pin7 to ground via a 50vr works perfectly for vgpu as far as i can tell.
start at 50k and decrease resistence to increase vgpu
core now does 650 with some artifacts in canyon and 625mhz artifact free.
all the +- 1.30volt measuring points in post#1 now read 1.38 idle and 1.40 when loaded.

i hope this is the right datasheet on the LM358: http://www.national.com/ds/LM/LM158.pdf
the RT9214 is usually pin6 to ground by what i can find on the web.
anybody got any tips on the memory mod before i proceed further.

here's a score at 650/700

and a bad pic of the mod :P

persivore · 05-09-2006, 07:57 AM

Connecting pin 6 to ground on the RT9214 via a VR approx 20x the value of the unmodded resistance to ground should work fine for the Vdd/Vmem mod.
If you want to check that the RT9214 is used for Vdd/Vmem, check the voltage at pin 8 on the RT9214. It should be approx 0.1 to 0.2v lower than the expected Vdd voltage supply

s e t h · 05-09-2006, 08:16 AM

awsome. thanks again.
with the system on i'm getting 1.85 from pin8 of RT9214.
with the system off pin#6 to ground reads 1.19Kohm
so multiply 1.2 by 20 and we get 24Kohm for the VR.
i have 20's and 50's with me so to be safe i guess i'll start with a 50 at max.
i was going to try a 50 off pin6 pretty soon anyway. thanks you for the confirmation.
knowing now that vdd is around 1.95 i'm wondering why this 1.2ns ram is so damn stubborn?
maybe it has a higher default voltage than the 1.4ns stuff at 1.8 (usually overvolted on 256mb 7900gt's to 2.0)

i wonder what LM358 has to do with all this? maybe vddq?

anyway i'll be back with results asap

humeyboy · 05-09-2006, 08:30 AM

The Samsung 1.2ns ram has a default voltage need of 2v and is rated for 800/1600ddr.

Im watching your mods closely to do my agp version of same spec card, so on the chip for GPU (LM358) what pin to ground to measure volts before I ground pin 7 using VR ?

Then same for Ram chip (RT9214), what point to ground to get proper reading of current memory voltage ?, then I can ground pin 6

alpha0ne · 05-09-2006, 08:30 AM

I think the 1.2 ram is rated to use 2.0V

humeyboy · 05-09-2006, 08:35 AM

Yes and I didnt make that up above, I got info from Samsung site for modules using the text on top of them.

alpha0ne · 05-09-2006, 08:49 AM

Some/all X1900XT's use the same ram and are known to use 2.0V

persivore · 05-09-2006, 08:57 AM

Accordig to this: http://www.samsung.com/Products/Semi...ily_cd=GME1002
There are 2 types of the memory which are rated to different voltages
-BJ - 2.0V+/-0.1V
-BC - 1.8V +/-0.1V

The stock voltage will depend on which suffix memory chips are used.

humeyboy · 05-09-2006, 09:29 AM

We have BJ=2v :-)

s e t h · 05-09-2006, 11:08 AM

hmmm....
ok i'm not sure RT9214 is vdd anymore.
i twisted that little vr till i was blue in the face
only to discover, apon checking, that i was connected to pin#7 instead of #6

*whew* at least the card survived

so reconnected properly to pin#6 and twisted till pin#8->ground reads 1.95volts.
ok this should now leave us with more vdd than these 1.2ns IC's are rated for.
"RESULTS" i hear you all scream

. well not much improvement on the memory at all.
not even from 740 to 760.

this leads me to believe that LM385 is the bugger we should be focussing on.

anyway over to you people for a few hours. i'm tired and don't wanna play anymore

good luck

edit: one other scary thing i noticed.
i'm driving a dfi-sli-dr with the sli jumpers in the SLi position instead of the single card position.
this leads to a lengthy period before the screen inits.
i noticed that, with the RT9214 mod in place at 1.95volts on pin#8, < this jumped all the way to 2.25volts during the pre-initialisation phase.
so lets all be carefull with RT9214 till some more experience has been gained.

humeyboy · 05-09-2006, 11:15 AM

Yes VR-Zone had a peep did mod to the chip your working on and got nothing for ram, its the other chip, was not same card but 2 same chips here:

http://forums.vr-zone.com/showthread.php?t=21846

LM324 (so L385 on yours ?) is one he ended up having to mod to get more volts to memory.

BTW there is a lot of volts flying about the card where you have measured, your core should only be 1.2v not 1.3v default so the points may not be the volts getting to core, ram needs 2v but until I get correct point to get true reading I wont know why it cant do 800/1600ddr stock, Im told its probably low like 1.7v or such.

s e t h · 05-09-2006, 10:54 PM

http://www.national.com/ds/LM/LM158.pdf
^ that is supposedly the datasheet on LM358.
i don't know if it's the right sheet but will someone please try confirm.

here's some readings from the ic with the system off.
pin2 = 250ohm
pin3 = 638ohm
pin4 = ground
pin5 = 638ohm
pin6 = 682ohm

and with the system running

pin1 = 3.9volts
pin7 = 5.5volts
pin8 = 12.24volts

if the datasheet is right then pin1 and pin7 are output A and B respectively.
other info on the web tends to indicate pin6 to ground is the usual method.
but based on the datasheet pin2 might be capable of varying the output on pin1.

so by my very limited estimation it looks like 20kvr->gnd from pin6 or 10kvr->gnd from pin2.

someone please try figure out whats going on before i kill something

thanks

btw pin2 and 6 both appear to be connected to a series of small components close to LM358.
is there anything i can do to help analize the situation to make the gueswork less?

graham_h · 05-09-2006, 11:11 PM

Keep up the great work.

s e t h · 05-09-2006, 11:21 PM

do you still have a card graham?
if so would you please take a very high detail shot of the LM358 and surrounding components.
as you can see the pictures in post#1 are from the photo's you posted previously

my camera sucks so i couldn'd possibly get a good enough photo to follow traces and such.
maybe it'll help some of the experts to identify what should be done next

graham_h · 05-09-2006, 11:23 PM

I'll pull my card out tomorrow and photograph it in bright sunlight.
Just don't stand a chance with a flash

Not sure how good it'll be though

humeyboy · 05-10-2006, 02:06 AM

I already have very high pics of both sides of card, the agp one but its 100% apart from sli part and hsi bridge, all the chips are same as yours.