6800le 'Black Screen' - The Complete Guide: Identification, modification and reason

Quote:

---

Originally Posted by ***Deimos***

Are you using default heatsink? How are the mem chips being cooled? I have a NV5 making firm contact with them. While running ATITool (with 5 min warmup) you can touch the back of the PCB where memory chips are. If ts hot, you need better cooling for memory. Its also a good idea to have at least a large fan blocking at the back where all the chips are in the corner. They get very very hot.

---

i have zalman fatality cooler and passive heatsinks from zalman for memory... on cpu i have titan vanessa s-type with 92mm fan that is placed vertical and it is blowing air to the other side of pcb... maybe the voltage is to low? i have gainward 6800le agp, and somebody on other forum said that gainward has memory at 2.55v, but i don't know what is the true? i can't measure it because i don't have equipment... i think that temperature shouldn't be a problem...

If you get 51xx without lod, that's actually very good :banana:
Pm me if you want some help beating me on the orb I won't mind being passed by ;)

2.2ns hynix needs 2.7v afaik, no way the voltage is that low. Did you do bios mod to increase vgpu? If not that corespeed is fantastic to :banana:
I can do around 420 on 1.1v, but up to 470 on 1.4v with zalman vf700cu. I also do have that fan blowing over pcb ( vreg on back off the card, caps near molex connector ) without that I get random blackouts. The ram get's warm to the touch after a bench on 970 and artifacts like hell even with 2.9v, score was even lower then with 950.

Quote:

---

Originally Posted by MartianInvader

If you get 51xx without lod, that's actually very good :banana:
Pm me if you want some help beating me on the orb I won't mind being passed by ;)

2.2ns hynix needs 2.7v afaik, no way the voltage is that low. Did you do bios mod to increase vgpu? If not that corespeed is fantastic to :banana:
I can do around 420 on 1.1v, but up to 470 on 1.4v with zalman vf700cu. I also do have that fan blowing over pcb ( vreg on back off the card, caps near molex connector ) without that I get random blackouts. The ram get's warm to the touch after a bench on 970 and artifacts like hell even with 2.9v, score was even lower then with 950.

---

i have 1.5v and 445mhz on core... can you post picture of your case near pcb?

Firechicken, jlovrek and MartianInvader ; Thank you for your support :toast:

Those replacement capacitors I have sent for severel users (~ 15 sets !!!) were obtained as samples from local supplier of Nippon Chemi-con. Thus, the most credits should be given to Chemi-con !! Once I got those I decided that the best purpose for them will be definately repairing of BS-cards. It was for me quite small task to send some capacitors, in the respect of PAIN of user having a BS-affected card without RMA possiblity !!!

Quote:

---

Originally Posted by ***Deimos***

I certainly hope thats with LOD and high performance in the driver. After about 380-400, my 6800 becomes very bandwidth limited, and thats with memory running at 900, 950, or even 972 (after voltmod). Go figure. With core only at 375, the GT runs memory at 1000, and Ultra with core at 400/425, memory at 1200.

for everybody who is only getting 800's with 2.2ns Hynix
get NIBITOR 2.8 from mkvtech
you can test timings in windows without flashing!!
for 900+, I found I MUST have tRCDRD=6 and tRP=6.

Intructions:
1. Tools-Read BIOS- Select Device-OK
2. Tools-Read BIOS- Read into NIBITOR- OK
3. Go to Timings tab, and click autoselect to change to proper timingset- ok
4. click detail timings, and you will see your timings.. make changes as desired. When you click ok, you will notice, that the Timing0, Timing1, Timing3 have changed. Write these down!
5. Back at the Timings tab, Click test timings, get timings, and insert the values you wrote down, click set timings, and go run atitool artifact tester (highly recommended) or some other benchmark.

EX:
Timing0: 060C1512
Timing1: 05010307
Timing2: 00240306

correspnd to:
tRC=18
tRFC=21
tRAS=12
tRP= 6
tRCDRD=6
tRCDWR=3
tRRD=4

---

Thanks for accurate timing tweak info !

One question releted to the test results you provided earlier. How did you got the 'local video memory usage statistics' active in RT ? I could not found those settings (or maybe I have too old version...)

Quote:

---

Originally Posted by jlovrek

i have zalman fatality cooler and passive heatsinks from zalman for memory... on cpu i have titan vanessa s-type with 92mm fan that is placed vertical and it is blowing air to the other side of pcb... maybe the voltage is to low? i have gainward 6800le agp, and somebody on other forum said that gainward has memory at 2.55v, but i don't know what is the true? i can't measure it because i don't have equipment... i think that temperature shouldn't be a problem...

---

The memory voltage of these cards is normally 2.72 V (+/- ~ 0.02 V). However, the speed of Hynix chips is set/limited by supply voltage and 2.95 is maximum acceptable value.


If you want to test whether increase of Vmem will have some effect there is an easy way to make mod:
http://img216.imageshack.us/img216/4...mod16ir.th.jpg

.. in theory it can be done without measuring actual voltage, BUT still I highly recommended measuring voltages before/after (and resistance of used resistor) !!

Quote:

---

Originally Posted by jlovrek

i have 1.5v and 445mhz on core... can you post picture of your case near pcb?

---

1.5v and 445 isn't great. Is that 1.5v flashed or 1.5v measured? I flashed 1.4 before and though I was getting that, but when I measured it it was still 1.1v. I flashed again at 1.6v and I now measured 1.4v. With 1.4v I can go to 470, but the scores are lower. If I had wc on the chip it would respond better I think :confused: because off less electro thermal dissapation or something like that.

I will put up pics off the pcb but I'm not sure what you would want to see?

Edit: the mod in the post above is truly easy, I done it like I saw it done by someone else with a fan connector molex so I can remove the mod when I'm not benching. I'm thinking next time I'll use a dipswitch though because getting a vr's legs to connect with the molex isn't allways 100% certain. I put small blobs off solder on the legs so I'm sure the contact is good, but that means I have to remember to power everything off because some tin could come loose and kill something. Which would have happend if I didn't have rememberd to power down lol, loose tin on mobo traces near some caps wasn't something I liked seeing :)

Quote:

---

Originally Posted by ***Deimos***

I certainly hope thats with LOD and high performance in the driver. After about 380-400, my 6800 becomes very bandwidth limited, and thats with memory running at 900, 950, or even 972 (after voltmod). Go figure. With core only at 375, the GT runs memory at 1000, and Ultra with core at 400/425, memory at 1200.

for everybody who is only getting 800's with 2.2ns Hynix
get NIBITOR 2.8 from mkvtech
you can test timings in windows without flashing!!
for 900+, I found I MUST have tRCDRD=6 and tRP=6.

Intructions:
1. Tools-Read BIOS- Select Device-OK
2. Tools-Read BIOS- Read into NIBITOR- OK
3. Go to Timings tab, and click autoselect to change to proper timingset- ok
4. click detail timings, and you will see your timings.. make changes as desired. When you click ok, you will notice, that the Timing0, Timing1, Timing3 have changed. Write these down!
5. Back at the Timings tab, Click test timings, get timings, and insert the values you wrote down, click set timings, and go run atitool artifact tester (highly recommended) or some other benchmark.

EX:
Timing0: 060C1512
Timing1: 05010307
Timing2: 00240306

correspnd to:
tRC=18
tRFC=21
tRAS=12
tRP= 6
tRCDRD=6
tRCDWR=3
tRRD=4

---

Also thanks for the timings info :toast: , never tinkerd with that before. Was hoping it will get me to 1gig ddr lol I know it has been done before but still it's a nice thing to archieve. Most untill now is 970 passing 05, and 03 always passed higher for me then 05 so I know I'm very close :)

For the Hynix memory chips, overvolting doesnt help at all.. just more heat, and more stability problems. But, you definetly need at least 2.7V. I found highest overclock at 2.8-2.84V: higher voltage was worse.

You definetly need a multimeter to measure the Vgpu and Vmem properly. Otherwise, you dont really know. The value in the VGA BIOS or Rivatuner (for Vgpu) is often wrong.

Strange mine did scale up to 2.9v, after that oc didn't increase. It is harder to keep stable but I don't blame that on my mem chips but my modding skills lol!

Wrong. It certainly wasnt your modding skills. Doesn't matter what temperature how well modded, there is a finite voltage limit, and a threshold beyond which you see sharp diminishing returns. Most ICs nowadays are made on very very fine lithography process with very small channels. When over-volting, you can quickly run into speed of light limit. Difference of voltage at the drain vs voltage at the source (ie, what you modded to) divided by the distance (typically 90-110nm) results in very high Volts/m gradient in the range of thousands and even millions of volts/meter gradient (velocity saturation and mobility degredation). Such highly energy electrons proportionally cause more damage to the silicon crystal lattice as hill:banana::banana::banana::banana: and other defects, shortening the life of the IC. Numerous other short channel effects also come into play like channel length modulation, subthreshold leakage, body effects, gate tunneling, and various temperature dependencies (carrier mobility decreases with increased temperature) to name a few.

Long story short, once you notice that increasing volts isn't helping much, its a sure sign that you have more than enough volts and that there are other limiting factors.

Quote:

---

Originally Posted by ***Deimos***

Wrong. It certainly wasnt your modding skills. Doesn't matter what temperature how well modded, there is a finite voltage limit, and a threshold beyond which you see sharp diminishing returns. Most ICs nowadays are made on very very fine lithography process with very small channels. When over-volting, you can quickly run into speed of light limit. Difference of voltage at the drain vs voltage at the source (ie, what you modded to) divided by the distance (typically 90-110nm) results in very high Volts/m gradient in the range of thousands and even millions of volts/meter gradient (velocity saturation and mobility degredation). Such highly energy electrons proportionally cause more damage to the silicon crystal lattice as hill:banana::banana::banana::banana: and other defects, shortening the life of the IC. Numerous other short channel effects also come into play like channel length modulation, subthreshold leakage, body effects, gate tunneling, and various temperature dependencies (carrier mobility decreases with increased temperature) to name a few.

Long story short, once you notice that increasing volts isn't helping much, its a sure sign that you have more than enough volts and that there are other limiting factors.

---

I agree most of that story, it is clearly seen you HAVE knowledge about submicron CMOS technology.

However, some comments: at first a short investigation about speed vs. voltage&temperature and at end more specific comment about this case.

In general the higher supply voltage the higher speed is obtained with a CMOS logic WHEN the junction temperature is kept constant. Self heating of circuit increases gate delays (decreases mobility inside channel), so in practice acquired speed is not linearily increased as a function of supply voltage.

I made a following electrical SPICE-simulation:

- ring-oscillator , 3 CMOS inverters
- Transistor parameters 0.13 um CMOS by IBM, used for GF6800/LE/GT chips
- Parameters from MOSIS (www.mosis.org)
- Transistor sizes PMOS: W = 2 um, L =0.13 um, NMOS W = 1 um, L =0.13 um (just an engineering guess..)

I simulated oscillation frequency as function of supply voltage at different temperatures. The frequency in the figure shown below is normalized (1 = 8.7 GHz).

http://img103.imageshack.us/img103/1...13ro2sz.th.jpg

Results:

1. Increasing the supply voltage increases the maximum operating frequency quite linearily (self heating is not considered)
2. Cooling is very important factor when a maximal operational speed is required.

Notes:

1. Accuracy of the models is not guaranteed
2. The minimum/maximum voltages (1.0 V, 1.6 V) might not be usable for GPU, since GPU is using much more complex logic.

---

Speed of Hynix-memories used in 6800-cards is of course more complex case than the investigation above. I think that those chips can operate in a reliable manner having supple voltage at least up to 2.95 V. This voltage level is specified in datasheet as a maximum operating voltage, and there is specified that 3.6 V as an absolute maximum voltage at any pin in any condition. Therefore, I suppose they have utilized a CMOS technology that can withstand such voltage level (2.95 V) without gate oxide degeneration (leakage etc.). Since, degeneration effects take place in long term I think those can not explain why speed is not increased anymore after some voltage level. An instantenous breakdown of e.g. gate oxide or source drain junction has of course catastrophic consecuenses and usually means a burn chip. But in short term as long as you keep below 'high voltages' and take care of cooling (=current/power does not 'melt' chip) I think the more voltage - the more speed should be expected. The same logic can be seen in datasheet, since they specify different supply voltages to chips with different speed specification (propably they are using the same IC-process for all versions and measure/select best out).

Another point is application of the logic, here the interface of the memory chips. The high supply increases I/O logic levels to/from GPU, and maybe after some level the I/O circuitry speed of GPU 'saturates'. Moreover, the observed speed of VGA-card is fuction of several different aspects and it is possible that after certain memory clock frequency some other bottle-neck will appear.

Quote:

---

Originally Posted by ***Deimos***

Wrong. It certainly wasnt your modding skills. Doesn't matter what temperature how well modded, there is a finite voltage limit, and a threshold beyond which you see sharp diminishing returns. Most ICs nowadays are made on very very fine lithography process with very small channels. When over-volting, you can quickly run into speed of light limit. Difference of voltage at the drain vs voltage at the source (ie, what you modded to) divided by the distance (typically 90-110nm) results in very high Volts/m gradient in the range of thousands and even millions of volts/meter gradient (velocity saturation and mobility degredation). Such highly energy electrons proportionally cause more damage to the silicon crystal lattice as hill:banana::banana::banana::banana: and other defects, shortening the life of the IC. Numerous other short channel effects also come into play like channel length modulation, subthreshold leakage, body effects, gate tunneling, and various temperature dependencies (carrier mobility decreases with increased temperature) to name a few.

Long story short, once you notice that increasing volts isn't helping much, its a sure sign that you have more than enough volts and that there are other limiting factors.

---

huh I thought I used a smiley in that post lol....
Guess I didn't. Point being I knew the short version, long version is way way way to xtreme for me but if my oc doesn't increase when upping the volts.... yeah I dig that :)

I dont know SikaRippa.. I always thought the capacitance was the major limitation for DRAM (like the Hynix 2.2ns GDDR sitting on my 6800). Afterall, as featuresizes get smaller, if you only had tiny area for the capacitor in each cell, you would need very fast refresh, or just wouldn't work. I believe they use various "roughing" techniques to increase the effective area of the capacitor. It has been a while so I'm not sure.

I would certainly be interested if somebody could confirm if large DRAM suppliers like Samsung use low-K dielectric between the interconnects to reduce interconnect to interconnect capacitance (it certainly seemed to be a big issue on complex R400 chips where the low-K really helped, but want to know if also used for high speed memory). As for the gate oxide, that can be improved with high-K dielectric. I dont think any range close to the supply voltage would be sufficient to cause any serious damage to the gate oxide. The kind of breakdown you descriibe I think is more typical of electro - static shock (typically hundreds to thousands of volts).

I am try give a short answer. The technology used for DRAM's is quite different to logic-CMOS procesesses, that's clear. The big issue with DRAM's is indeed: how-to-make-as-much-as-possible-capacitance-in-a-very-tiny-area. All methods are used as you already listed, and the CMOS part can be quite poor compared to state-of-art CMOS. However, once technology (=capacitance value/square um) is fixed and memory cells are designed & laid out the speed of a memory cell vs. supply voltage should have similar characteristics as in standard CMOS i.e. the higher supply voltage the higher possible access rate. The capacitance keeping DRAM state should remain quite constant while supply is higher level, so it does not have influence to speed. Another point, why Hynix specify higher supply voltage for higher frequencies, if supply does not have effect to speed ?

Breakdown voltages of submicron CMOS-transistors are usually in magnitude of ~2 times of supply voltages. Generally the weakest part is gate oxide, which starts to tunneling heavily ~ 2*VDD voltage level ((yes, I have measured myself)). The source and drain connections can handle higher voltages.

---

Summary (IMHO):

In theory: higher supply voltage gives higher speed, when temperature is low enough and supply voltage is lower that absolute maximum ratings.

In practice: The obtained speed in a high APPLICATION LEVEL (like running 3d benchmark) is sum of several factors: GPU/memory interface, temperature of all components, software etc. As a result some users do not get significant improvement to 'memory frequency' while having higher voltages.

Oh I am not arguing at all that higher supply voltage will increase the speed of the DRAM. I can confirm that easily with my BH5 memory! However, I am saying that there are voltage threshold beyond which adding more voltage will not help increase Mhz, regardless of cooling etc, and that this voltage threshold is dependent on the manufacturing process and design of the IC (some are more tolerant like famous Winbond BH5)

ok, today i have replaced zalman fatality cooler with nv5 silencer in hope to get +900mhz on my 2.2ns memory... my temp. went down for about 5-10degrees at load, but my memory isn't more overclockable then before... i'm so :mad: !!!! it stil at 860mhz without artifacts! what should i do? voltmod? hm, i still can't measure my voltage of memory... sikarippa, with your caps, is there any change in voltage of memory? cause they have higher capacity...

Quote:

---

Originally Posted by ***Deimos***

Oh I am not arguing at all that higher supply voltage will increase the speed of the DRAM. I can confirm that easily with my BH5 memory! However, I am saying that there are voltage threshold beyond which adding more voltage will not help increase Mhz, regardless of cooling etc, and that this voltage threshold is dependent on the manufacturing process and design of the IC (some are more tolerant like famous Winbond BH5)

---

I think the case about Vmem & OC-capability is now fully considered. In the case if there was (?) a 6800-card having separed VDD/VDDQ voltages some kind of extra tweaking/testing could be possible by changing the VDD/VDDQ -ratio.

Quote:

---

Originally Posted by jlovrek

ok, today i have replaced zalman fatality cooler with nv5 silencer in hope to get +900mhz on my 2.2ns memory... my temp. went down for about 5-10degrees at load, but my memory isn't more overclockable then before... i'm so :mad: !!!! it stil at 860mhz without artifacts! what should i do? voltmod? hm, i still can't measure my voltage of memory... sikarippa, with your caps, is there any change in voltage of memory? cause they have higher capacity...

---

Voltage mod might give some boost and the replaced capacitors should also enable better OC-capability. However, if the solderering of the capacitors is not proper (=cold solder) this has a huge effec: how well you soldering proceeded ?

It is not suggested to make Vmod's without measuring. The mod I proposed is very easy to do even without mesuring, but allways something can be happened and it that case you might have running card (for a while) and too high voltage levels (short life).

Quote:

---

Originally Posted by SikaRippa

Voltage mod might give some boost and the replaced capacitors should also enable better OC-capability. However, if the solderering of the capacitors is not proper (=cold solder) this has a huge effec: how well you soldering proceeded ?

It is not suggested to make Vmod's without measuring. The mod I proposed is very easy to do even without mesuring, but allways something can be happened and it that case you might have running card (for a while) and too high voltage levels (short life).

---

well, the soldering has passed very well, that was doing a tv-service man... i think that solder of caps isn't a problem cause everything works great, i don't have bs any more, just my memories couldn't pass 900mhz withous artifacts... here are numbers from my memory chips:
HY5DU283222
AF-22 338A

Quote:

---

Originally Posted by jlovrek

well, the soldering has passed very well, that was doing a tv-service man... i think that solder of caps isn't a problem cause everything works great, i don't have bs any more, just my memories couldn't pass 900mhz withous artifacts... here are numbers from my memory chips:
HY5DU283222
AF-22 338A

---

OK, soldering should not be problem. The point of my question was, that users have been complaining not to get good performance (or remove BS) even when capacitors replaced. In many cases (AFAIK) the problem was due to cold soldering which increases series resistance = similar effect than ruined capacitors.

Your memories should be capable to run >900 MHz, so Vmod might help (900 MHz operation requires 2.8 V instead of 2.72 V according to datasheet).

jlovrek... if you only use Vmem mod and capacitor mod you shouldn't expect more than 900Mhz. You need to change the timings. I have posted detailed instructions and several sets of timings to try in previous posts here.

Quote:

---

Originally Posted by ***Deimos***

jlovrek... if you only use Vmem mod and capacitor mod you shouldn't expect more than 900Mhz. You need to change the timings. I have posted detailed instructions and several sets of timings to try in previous posts here.

---

I think that luck (=memory chips are unique) plays important role: some users have managed to get ~ 1000 MHz even at stock voltages (and even more with higher voltage). Take a look post #11:

http://www.xtremesystems.org/forums/...ad.php?t=62651

Most 6800 nu/le have Hynix GDDR (typically 2.8ns or 2.2ns). 2.8ns~DDR700, 2.2ns~DDR900.
I got the biggest boost in memory speed on my card by changing the timings. Before I could barely get up to 900. Afterwards, 960+ (60Mhz increase just by chnaging tRP=6, and tRCDRD=6). With Vmem almost 984 stable.

Quote:

---

Originally Posted by ***Deimos***

Most 6800 nu/le have Hynix GDDR (typically 2.8ns or 2.2ns). 2.8ns~DDR700, 2.2ns~DDR900.
I got the biggest boost in memory speed on my card by changing the timings. Before I could barely get up to 900. Afterwards, 960+ (60Mhz increase just by chnaging tRP=6, and tRCDRD=6). With Vmem almost 984 stable.

---

i have tried with all sets of timings and still nothing... damn!