vranoga1
08-25-2004, 04:38 AM
I have 3 pieces of GeCUBEs 9550XT 128mb
This is Radeon 9550 chipset, however they are built on 9600XT PCB with 3.3ns BGA memory.
Standard frequency is 400core, 250(500) memory.
Without voltmods they all exceed 500Mhz core (with better cooling; 505-520)
On all three; memory maxes out exactly on 288Mhz (576ddr).
It is a HECK of a deal for a 70eur card (cca 75USD)!!!!! (when overclocked, they are faster then standard 9600XT)
now the problem:
I figured out the voltmodes and applied them to all three cards.The problem is the memory clock;
whenever I raise any voltage mem overclock seems to drop. For raising booth VCore and VMem.
Here are excact results for 2 cards for Vmem mods:
-----------CARD 1)------------
______________________________
VGPU standard
---------
Ustand.=1.30 V
Rstand.=240 Ohm
Fmax stand.=504Mhz
VMem standard
---------
Ustand.=2.89 V
Rstand.=671 Ohm
Fmax stand.=288Mhz
VMem mod1
---------
Umod.=2.94 V
Rmod.=659 Ohm
Fmax mod.=281Mhz
VMem mod2
---------
Umod.=3.01 V
Rmod.=642 Ohm
Fmax mod.=264Mhz
-----------CARD 2)------------
______________________________
VGPU standard
---------
Ustand.=1.30 V
Rstand.=252 Ohm
Fmax stand.=517Mhz
VMem standard
---------
Ustand.=2.90 V
Rstand.=677 Ohm
Fmax stand.=288Mhz
VMem mod1
---------
Umod.=2.98 V
Rmod.=652 Ohm
Fmax mod.=270Mhz
VMem mod2
---------
Umod.=3.14 V
Rmod.=612 Ohm
Fmax mod.=unobtainable. Artifacts at any frequency
------card 3 is pretty much the same. (same trend in voltage/clock)
__________________________________________________ __________________
-so, conclusion is; raising memory voltage, lineary decreases mem clocks! Ok, lets go further...
-more unusual is that raising core voltage also decreases mem clocks. It also drops lineary, but the drops are not
as rapid and big as with mem voltage (5-20mhz drop with increased VCore vs. 10-50mhz with increased mem voltage).
-On the other hand core frequency gains are normal, resulting in 540-590mhz core frequency with 0.1-0.2V VCore
increase.
THE BIG QUESTION: why does mem overclock always drop with any increased voltage?
AND: What should be done to break that trend? OR; how to increase mem clocks?
(I also have to mention: heat is not the issue (the much better cooling is used). Neither on the Core, neither on
mem..)
DETAILS:
-Voltage regulation chip is: "1575 PFEQ". Pin 1 is the ground. Pin 2 is the feedback pin.
-memory is Etrontech BGA 3.3ns. So it is rated at min. 300Mhz. PN is EM6A9320BI-3.3
it is rated for CAS3, CAS4 and CAS5 operation. I eliminated the timing problem. With RaBit 1.3 I've edited ram
timings and tried many combinations (CAS3, CAS4 and CAS5 included). I also copied timings from many others 9600XT
bioses (mostly GeCubes). It did'nt help at all; standard timings showed highest stability/performance/clocks.
-I also observed two, maybe important details:
1) voltage regulation circuits are exactly the same for both core and memory! (same controler chip, same two
mosfets, the same four capacitors and the same coil..)
2) Standard memory voltage is unusualy high: 2.90V (it also doesn't tolerate more than 3.1V)
MY BEST GUESES ABOUT THIS PROBLEM:
1) vdd and vddq are not in optimal correlation for this PCB/memory combination!?!
I've read that there is no universal formula for vdd/vddq correlation. In theory vddq should be slightly lower or
equal to vdd. But in practice, it varries alot. Some manufacturers are using half vdd voltage for vddq, and some
are even using HIGHER vddq than vdd. By the Entrontech PDF file, this memory operating conditions are 2.8V for VDD,
and 2.8V for VDDQ
2) there is some kind of excess voltage/power drain somewhere on the card when raising VCore/VMem!?!
---------------------------------------------------------------------------------------------------------
So, what do you think?
PICTURES ARE ATTACHED..
Also, I would appreciate some help with vddq mod! The VCore and Vmem mode were simple (they were in fact the same),
but I don't have a clue where to start with the vddq mod? Where to measure? how to raise it?
For now I have a feeling that the vddq is the main guilty factor; and that vddq mod is the answer...
Regards to all..
http://web.vip.hr/mrgoran.vip/gecube-9550u-back-with-marks.jpg
http://web.vip.hr/mrgoran.vip/gecube-9550u-front-with-marks.jpg
This is Radeon 9550 chipset, however they are built on 9600XT PCB with 3.3ns BGA memory.
Standard frequency is 400core, 250(500) memory.
Without voltmods they all exceed 500Mhz core (with better cooling; 505-520)
On all three; memory maxes out exactly on 288Mhz (576ddr).
It is a HECK of a deal for a 70eur card (cca 75USD)!!!!! (when overclocked, they are faster then standard 9600XT)
now the problem:
I figured out the voltmodes and applied them to all three cards.The problem is the memory clock;
whenever I raise any voltage mem overclock seems to drop. For raising booth VCore and VMem.
Here are excact results for 2 cards for Vmem mods:
-----------CARD 1)------------
______________________________
VGPU standard
---------
Ustand.=1.30 V
Rstand.=240 Ohm
Fmax stand.=504Mhz
VMem standard
---------
Ustand.=2.89 V
Rstand.=671 Ohm
Fmax stand.=288Mhz
VMem mod1
---------
Umod.=2.94 V
Rmod.=659 Ohm
Fmax mod.=281Mhz
VMem mod2
---------
Umod.=3.01 V
Rmod.=642 Ohm
Fmax mod.=264Mhz
-----------CARD 2)------------
______________________________
VGPU standard
---------
Ustand.=1.30 V
Rstand.=252 Ohm
Fmax stand.=517Mhz
VMem standard
---------
Ustand.=2.90 V
Rstand.=677 Ohm
Fmax stand.=288Mhz
VMem mod1
---------
Umod.=2.98 V
Rmod.=652 Ohm
Fmax mod.=270Mhz
VMem mod2
---------
Umod.=3.14 V
Rmod.=612 Ohm
Fmax mod.=unobtainable. Artifacts at any frequency
------card 3 is pretty much the same. (same trend in voltage/clock)
__________________________________________________ __________________
-so, conclusion is; raising memory voltage, lineary decreases mem clocks! Ok, lets go further...
-more unusual is that raising core voltage also decreases mem clocks. It also drops lineary, but the drops are not
as rapid and big as with mem voltage (5-20mhz drop with increased VCore vs. 10-50mhz with increased mem voltage).
-On the other hand core frequency gains are normal, resulting in 540-590mhz core frequency with 0.1-0.2V VCore
increase.
THE BIG QUESTION: why does mem overclock always drop with any increased voltage?
AND: What should be done to break that trend? OR; how to increase mem clocks?
(I also have to mention: heat is not the issue (the much better cooling is used). Neither on the Core, neither on
mem..)
DETAILS:
-Voltage regulation chip is: "1575 PFEQ". Pin 1 is the ground. Pin 2 is the feedback pin.
-memory is Etrontech BGA 3.3ns. So it is rated at min. 300Mhz. PN is EM6A9320BI-3.3
it is rated for CAS3, CAS4 and CAS5 operation. I eliminated the timing problem. With RaBit 1.3 I've edited ram
timings and tried many combinations (CAS3, CAS4 and CAS5 included). I also copied timings from many others 9600XT
bioses (mostly GeCubes). It did'nt help at all; standard timings showed highest stability/performance/clocks.
-I also observed two, maybe important details:
1) voltage regulation circuits are exactly the same for both core and memory! (same controler chip, same two
mosfets, the same four capacitors and the same coil..)
2) Standard memory voltage is unusualy high: 2.90V (it also doesn't tolerate more than 3.1V)
MY BEST GUESES ABOUT THIS PROBLEM:
1) vdd and vddq are not in optimal correlation for this PCB/memory combination!?!
I've read that there is no universal formula for vdd/vddq correlation. In theory vddq should be slightly lower or
equal to vdd. But in practice, it varries alot. Some manufacturers are using half vdd voltage for vddq, and some
are even using HIGHER vddq than vdd. By the Entrontech PDF file, this memory operating conditions are 2.8V for VDD,
and 2.8V for VDDQ
2) there is some kind of excess voltage/power drain somewhere on the card when raising VCore/VMem!?!
---------------------------------------------------------------------------------------------------------
So, what do you think?
PICTURES ARE ATTACHED..
Also, I would appreciate some help with vddq mod! The VCore and Vmem mode were simple (they were in fact the same),
but I don't have a clue where to start with the vddq mod? Where to measure? how to raise it?
For now I have a feeling that the vddq is the main guilty factor; and that vddq mod is the answer...
Regards to all..
http://web.vip.hr/mrgoran.vip/gecube-9550u-back-with-marks.jpg
http://web.vip.hr/mrgoran.vip/gecube-9550u-front-with-marks.jpg