7800GT new layout, hardware modding

[t024484]

After having used many ATI boards, from 9500 to X850XT PE, I recently switched to a 7800GT board because on average it is better and cheaper than the new X1800XL.

The one I bought was a XFX, code PV-T70G-UDE3
450/1050 3dmark05 7221

Tuning it via Coolbits, brought me to:
462/1100 3dmark05 7372

Looking for more, I tried to change the firmware, with and without using the 6800 saving trick.
Although I could indicate that the 3D voltage should be 1.5 volt instead of 1.4 volt, the result was worse than before, and measuring the Voltage on my board showed me that the 1.4 Volt did not change at all, so something changed but not the 1.4 Volt.

Then I decided to do some hardware modding.
The problem though was that the layout of my board was completely different to pictures in previous postings by Shamino and others.

I made pictures from the left lower and the left upper side (as seen from the back)

41841

41842


I did not want to have wires and potmeters on my board, so I made the changes with SMD resistors, like the ones already sitting on the board.

The first picture is marked red on the spot where you can find a resistor with the coding 01B, in this case meaning 1Kohm. Lowering the value of this resistor increases the value of the memory voltage. I soldered in parallel a resistor of 15Kohm. This raised the memory voltage from 2.14 to 2.23 Volt.
To measure this voltage, go upwards in the picture where you see two fat round solder points, coming from the wires of a coil on the other side.
The right one of these two gives you the Mem. Voltage.

The red marking in the second picture is where you will also find a resistor with the coding 01B, or 1Kohm. I soldered in parallel a resistor of 10Kohm, raising the CPU voltage from 1.42 to 1.51 Volt. You can measure this voltage on the big round solder point almost directly above and slightly to the left of the red marking.

Of course will it be possible to use normal resistors and/or trim potmeters instead.
I just liked to have a board that looks unmodified.

RESULTS

I tried with a geometric delta clock of 60, 40 and 0 Mhz

With 0 Mhz the maximum achievable clock was 481/1200 3Dmark05 7638
With 40 Mhz I could get 472/1200 3Dmark05 7731
With 60 Mhz also 472/1200 3Dmark05 7763

The problem that I had finding the max clock was in the second game of 3Dmark05.
This game, firefly forest, is extremely more sensible than the other two games to a clock that is too high. I have no idea why. It first starts with producing artifacts and shortly after the system hangs, almost independent of GPU temp.
ATI Tool advised higher frequencies, but firefly did not accept these.

One last remark concerning my board: Although it looks in the firmware as if you can program the speed of the ventilator, the ventilator has 5.1 volt on its pins all the time, regardless of 2D, 3D or firmware setting. The GPU remained below 70C all the time with this standard cooler.

It was fun doing, and I have a board with comparable speeds to a 7800GTX board.

P5AD2-Premium
1Gbyte OCZ 6400EB
P4 3.46EE@3.8Ghz


41843

[Raphie]

Have exactly the same card 8735 (http://service.futuremark.com/compare?3dm05=1625695) @ 529 / 1242 unmodded..... BTW GTX boards go around 9000+

[Supertim0r]

955XE ?
is it the Black XFX 7800GT
i've the regular 7800GT (450/1050) running @ 500/1200...vmods SOON :)

[teko]

I'm doing 490/1180 so far on my black PCB XFX stock, so it's top notch that you have found the needed point's to do the hardmod on this PCB :) Even though I might have to wait a little until I'm financialy able to take the chance :P

[OV3RCLOCK3R]

my club3d 7800gt does 510 core and 1.22ghz ram on auto detect:D

[t024484]

My unmodified XFX did 499 and 1.13 on auto detect, but 3Dmark05 dit not allow for more than 462 and 1.10.
Auto detect brings my GPU temp only to 51C, while running 3Dmark05 shows me a GPU temp of max 70C.
So I think that this large temp difference causes the discrepancy.
I will replace the standard cooler for a Zalman VF-700 and see what happens.


P.S. nevertheless you are a lucky bastard with your very high 510/1.22

[OV3RCLOCK3R]

My unmodified XFX did 499 and 1.13 on auto detect, but 3Dmark05 dit not allow for more than 462 and 1.10.
Auto detect brings my GPU temp only to 51C, while running 3Dmark05 shows me a GPU temp of max 70C.
So I think that this large temp difference causes the discrepancy.
I will replace the standard cooler for a Zalman VF-700 and see what happens.


P.S. nevertheless you are a lucky bastard with your very high 510/1.22

hehehhe, lol.

505/1.22ghz is fully stable, no artifacts or anything, 8.5k in 3dmark05 and 18.5k in 3dmark03:D:D

[Postal Dude]

my card is very similar overclocker. Auto detects 523/1240. Can run 510/1220 without artifacts. If i go upto 515 i get minor artifacts. Still pushin the memory. I have a standard Leadtek GT. Some sort of special edition standard one tho. It has a HS on the rear of the PCB with thermal transfer pads on the rear of the gddr and core. Dunno how effective it is but it looks good

[t024484]

Today I installed a Zalman VF-700 cooler instead of the standard cooler.
The temp under full load went down an amazing 15C.
The max achievable clock with 3Dmark05 with the better cooling, could not be raised at all. Very disappointing.

Whith the cooler removed, I could see the memory chips, Samsung K4J55323QF-GC20.
I googled the specs, and found that XFX is using a timing that is below the Samsung specs.
Just to see what happened, I changed the timing in the firmware to the Samsung timing.

-------------XFX---Samsung
Tras---------13-------15
Trc----------19-------21
Trfc---------24-------27
Trcdrd--------8--------8
Trcdwr-------3--------4
Trp-----------5--------6
Trrd----------4--------5

The max artifact free mem clock with ATI Tool went up from 1200 to 1370.

I tried 3Dmark05 to test this higher mem clock:
Before with 472/1200 3Dmark05 gave me 7763

Now with 472/1350 I got 7892, quite an improvement for a memory timing that is worse than before!

Since the temp was so much lower with the new cooler, I changed the 10Kohm for a 6K8 resistor, to give me a CPU voltage of 1.55Volt.

What is quite frustrating now, is that Nvidia detects a max frequency of 505/1150 and with ATI Tool I have a totally artefact free maximum at 486/1370, but although in the “Battle of Proxicon” everything goes fine with these high clocks, it is impossible as before to raise the GPU clock beyond 472 in the “Firefly Forrest”.
The system hangs immediately after the start of this second game.

But with 472/1370 everything is completely stable.

41875

PS because of the geometric delta clock of 50 that I am using, Futuremarkk shows a clock of 522/1350,
So with all the mods I have made, these are the max clocks that I can achieve.

http://service.futuremark.com/servlet/Index?pageid=/orb/projectdetails&projectType=12&projectId=1661106

[OV3RCLOCK3R]

my card is very similar overclocker. Auto detects 523/1240. Can run 510/1220 without artifacts. If i go upto 515 i get minor artifacts. Still pushin the memory. I have a standard Leadtek GT. Some sort of special edition standard one tho. It has a HS on the rear of the PCB with thermal transfer pads on the rear of the gddr and core. Dunno how effective it is but it looks good


What ram has your card got on it?

Mines the Infineon 2.0ns rated for 1.0ghz but its still doing over 1.22ghz:D

Any pics of the card, wanna see what you are talking about with the HS on the back:)

[Postal Dude]

yo overclocker. I iant had a chance to take the HSF off yet but will look as soon as i can. search for a thread made by me called 'to anyone with a 7800gt' There are pictures of the rear of my card in there. Found out that it does actually take away quite alot of heat. Gets nicely toasty during rthdribl

[BadlyDrawnBoy]

I've got one of the new ones. Running it at 475/1200 now. It's in a QPack, so I don't want to push it too far. Have the NV Silencer on it too because the stock fan is pretty loud.

[teko]

@ t024484 Could you try to run some gaming benchmark and see what the difference is in FPS?

[bAdboLL]

I got the XFX 7800GT with a blue pcb.

Does this work on my card also?
I dont think so =/

Here is a picture of my card;
http://hem.bredband.net/badboll/7800gtvoltagemod.jpg

Have i marked the right areas on the board? It doesnt look like your pics =/

[SynGroW]

I have 7 XFX 7800GT's here. 4 green w/green pcb, 2 w/black pcb and a blue one.

The blue card does about 520/1200 3DMark stable
The black cards do about 490/1250 3DMark stable

Still have to test all the green ones :p

Adding better cooling gave me about 10MHz more GPU (TTBT instead of stock cooling)

I was wondering if anyone can post a bios with the 1.5v mod and the tweaked timings to use in these cards :)

Gonna test the green cards today, I'll post some pics of the three different cards this evening.

[bAdboLL]

I have a bios with 1.5V, but not changed any timings. Dont know how do do that either, anyone that have som nice GT memory timings they want to share?

Here is my XFX 7800GT 1.5V bios anyway.

http://hem.bredband.net/badboll/gtny.rom

[Kasparz]

I got one 7800gt now 1.5Vcore/2.21Vdd.
Stable at 500/1430, but getting very hot with stock HS.
I got 8.7k 05 marks on Venice@2.5Ghz, so i install new OS, and try to reach 9k tomorrow.

[t024484]

I have a bios with 1.5V, but not changed any timings. Dont know how do do that either, anyone that have som nice GT memory timings they want to share?

Here is my XFX 7800GT 1.5V bios anyway.

http://hem.bredband.net/badboll/gtny.rom

Have you measured if you really had 1.5 Volt after the bios mod?
Your bios looks exactly as mine after the addition of the 1.5Volt.
The effect was everything else but not an increase from 1.4 to 1.5.
I would be amazed if it worked for you.

Looking at your timing tables, it is quite obvious that you have a different ram. If you tell me what ram you have, I can probably tell you the timing that you need, compared to what you have.
You get acces to the memory if you take of the cooler.
You cannot just use somebodies else's timing. You have to now the part number of your ram.
Changing the timing is very easy with Nibitor. Nibitor will tell you which of the 7 tables is used for your memory.

[saaya]

welcome to XtremeSystems t024484 :toast:

what geometric clock do you mean?
the reference clock for the vpu?

and why did you use smd resistors?
use a pot first and then find out what the best vgpu/resistance is and replace the pot with an smd resistor, but now you will have to solder on your card again and again until you find the sweet spot :D

[bAdboLL]

Before the bios mod i could OC my card to 463/1230, and after i could run stable up to 495/1250.
So surely i have gain more voltage, no doubt :)

Infineon HYB18T256324F-20, thats what the text on my mem's say.

Hm, so really, could i gain any performance by just changing my timings? Sounds good :)

[t024484]

welcome to XtremeSystems t024484 :toast:

what geometric clock do you mean?
the reference clock for the vpu?

and why did you use smd resistors?
use a pot first and then find out what the best vgpu/resistance is and replace the pot with an smd resistor, but now you will have to solder on your card again and again until you find the sweet spot :D

The geometric delta clock is added to the clock that you specify (root clock) to give you the Vertex Clock.
So with a 472 as an input, my real Vertex clock is 522Mhz, and my ROP/Shader clock is 459Mhz ( 3dmark05--7895)
With a geometric delta of 0, all clocks are 472 when the input is 472. I could go to a max clock input of 483 with delta 0, but this gave me an inferior score (3Dmark05--7660).

Most 7800GTX boards have this clock set at 40 Mhz.
In my case on my XFX 7800GT it was set on 0 Mhz.

Look also here
http://www.xtremesystems.org/forums/showthread.php?t=72638&highlight=7800gt

I did not have to test and try many SMD resistors, one was enough.
Connected to the 1.4Volt output are two resistors in series, acting as a voltage divider.
The sensing point between the two is regulated to be exactly at 0.5 Volt.
Since the resistor connected to ground is 1Kohm, it is very easy to calculate that in the case of the GPU the second resistor is 1.8Kom, to give you 1.4 Volt.
Just as simple is it to calculate that 10Kohm in parallel to the 1 Kohm resistor raises the output to 1.5Volt, and with 6K8 ohm to 1.55 Volt.

For the memory the calculation is similar. Also 1 Kohm to ground and also 0.5Volt reference, but here the output went from 2.1 to 2.2Volt.

So no trial and error, but directly spot on.

[RaZz!]

damn, i also want a 7800gt which is getting over 500/1200 with the auto detecion of the forceware lol :(

with auto detect i get 465/1130 .... artifact-free in all games only with 440mhz on the core (with a delta of 40). the maximum clock for the mem is 1180. anything higher will cause systemfreezes.

i'm totally unhappy with that crap overclocks on my gt... :( :'(

@threadstarter: what forceware are you using? my gt @440/1100mhz with a 4000+@2.4ghz scores 7555 3dmarks in 3dmark05 (forceware 81.98)
strange that urs only scores ~7300 with over 460 mhz one the core O_o

[bAdboLL]

The strange thing is that i could pass 3dmark05 with a coreclock of 526mhz, that was with delta 40.
With delta 0 my max stable is 495mhz.

t024484, could you please check my memorytimings up? And how much can it gain?

[t024484]

The strange thing is that i could pass 3dmark05 with a coreclock of 526mhz, that was with delta 40.
With delta 0 my max stable is 495mhz.

t024484, could you please check my memorytimings up? And how much can it gain?

I do not know which timing table you are using, you can get this info through Nibitor with "autoselect timingset".
Looking at your Bios, I think most likely it will be set 5.
I cannot tell you if and what the difference will be, i can only tell you what the optimal timing according to Infineon, the supplier of your ram chips, should be.

Suppose you have timing table 5:

Yours---Infineon
19--------19
27--------27
12--------12
7----------7
8----------8
5----------5
4----------4

So if I you have set 5, your table is spot on, and cannot be improved.
How much worse was my table.
Let me now if timingset 5 is the valid one in your system..

[t024484]

The strange thing is that i could pass 3dmark05 with a coreclock of 526mhz, that was with delta 40.
With delta 0 my max stable is 495mhz.

t024484, could you please check my memorytimings up? And how much can it gain?

If you could use 526 with a delta of 40, you probably mean that you specified a clock of 486Mhz, which is very good.
I achieved 522 with a delta of 50 with 1.5Volt, but exactly at that point if you increase by 1 Mhhz, the ROP/ Shader clock jumps from 459 to 486 Mhz, causing a major step in performance.

The next 27Mhz step for the ROP/Shader is at 500Mhz, and until there the ROP/Shader stays ar 486Mhz.

Obviously your GPU could handle this 486 Mhz, and mine couldn't.
With the Delta of 0 you achieved 495, which is also the ROP/Shader clock, and is slightly more than the 486 with the delta 40.
But your Vertex clock is 526 with delta 40 and only 495 with delta 0, so probably the delta 40 gives you the best results.

[bAdboLL]

I think im going to try delta 20-25 later, do you think thats better?

And btw, i got ALOT of artifacts at 486 + 40, but atleast i could run it without crash.
8165p.

But i would really like to play with some timings and deltas later, 8300 here i come! ;)

[bAdboLL]

"Your Timingset is Nr5"

So there is nothing i can do about the timings then?

[t024484]

"Your Timingset is Nr5"

So there is nothing i can do about the timings then?

I do not know if your set is nr 5. Check it with Nibitor.

[bAdboLL]

Yeah i did that and Nibitor said ""Your Timingset is Nr5" and the numbers where exactly like the one you posted here before.

[t024484]

Yeah i did that and Nibitor said ""Your Timingset is Nr5" and the numbers where exactly like the one you posted here before.

The way left to get your memory faster, is to Volt mod it from 2.1 to 2,2
You could also try a more relaxed mem timing.
Infineon specifies for 1200Mhz
23
33
15
8
10
7
5

But you could also try something in between like
21
30
14
8
9
7
5

Good luck

[bAdboLL]

What will happend if i try more relaxed timings?

Will i be able to clock higher?..

[DataNusse]

What will happend if i try more relaxed timings?

Will i be able to clock higher?..
Yeah thats right. Memory clocks more.


This sound really nice, so by playing with Delta temp you can get better scores? This i need to tray out on my GT.

[bAdboLL]

I just tried delta 20, and it works perfectly!

I can now clock it to 480/1250 (501/1250) totaly stable.
Before when i had delta 0 i could just clock it to 485/1250..

With the new values i get almost exatly 8010p in 05, pretty ok i think!

I will also try to relax the mem timings later on, maby i can get the mems up to 1300mhz+ without vmod.

[t024484]

@ t024484 Could you try to run some gaming benchmark and see what the difference is in FPS?

Teko, I already mentioned this in my postings, but here they are again:

Sandard XFX7800GT Black version 450/1050
3dMark05 score: 7221

Tuned XFX7800GT Black version 522/1350
Delta of 50Mhz (clock displayed in Coolbits is 472Mhz)
Vgpu modded from 1.4 to 1.5 Volt
Vmem modded from 2.1 to 2.2Volt
Mem timing adjusted to Samsung specs
3dmark05 score: 7892

So expect an increase in FPS of roughly 10%

Look also at http://www.xtremesystems.org/forums/showthread.php?t=88069

I raised the score to 8114 by changing a current limitor and increasing the 1.5Volt to 1.5Volt.

Increase is now almost 13%