End the Johnny Guru Rumor/8800GT Memory Interface Issue?

bildo123

Distinguished
Feb 6, 2007
1,599
0
19,810
Ok, whats the deal, there's a lot of guys in this forum that ahve knowelege beyond their years and I want to summon them to end this damned rumor IF it's a rumor...I'll link to the single post...

http://www.hardforum.com/showpost.php?p=1031607668&postcount=13

Ok, so is this BS or is this really true? On stock settings my memory on my BFG 8800GT OC can hit 1040 rock stable. Not only that I ran some tests and going from 900 to 1040 provided a 6FPS increase in Portal with maxed out settings at 1920x1200. I think thats amazing, probably less for other games but it's the principle... Does anyone know of, or has anyone run their own 8800GT for months on end at a memory clock at or above 1Ghz?
 

homerdog

Distinguished
Apr 16, 2007
1,700
0
19,780
This was only recently explained to me, but I'll give it a shot. The wire traces of the 8800 series' memory bus are all pretty close together. This causes an excessive amount of "noise" between the traces, which is not good if you are trying to achieve high frequencies. This is why Nvidia can't use GDDR4 on any of their current cards. It may also be the reason that it is so hard to OC the memory on the 8800GT. Mine won't go a MHz past 950 without becoming unstable.
 

sailer

Splendid
Homerdog, you're pretty close on your explaination. Yes, there is excess noise due to the materials that Nvidia is using on its board. Along with the noise, as I understand it, there is the problem of electromigration, which is what jonnyguru is talking about. Noise is troublesome, but doesn't cause damage in and of itself. Its kind of like hearing a lot of static when listening to a radio.

Electromigration is when the physical properties of the chip (the wires, transistors and such), are being changed because of the heat and the electricity going through them, and parts of the chip are disintegrating and moving to other parts of the chip. Try thinking of it this way; you put a drop of blue paint on a piece of paper and a drop of yellow paint next to it. They are two separate colors. Now mix the drops. You end up with green paint. The same idea happens with electromigration. At higher clocks, using more voltage (electricity) and thus causing more heat, some of one wire will move to another wire, thus changing them both. This is irrepairable damage. At first, it may not be noticeable, but then suddenly the card quits working as a wire either breaks apart or gains too much material to work properly, perhaps even fusing with an adjacent part. Some of the problem of electromigration can be abated by using better GPU coolers to lower the heat and thus stave off the electromigration, same as the reason we use CPU coolers. But, with time, the damage will still occur. It just takes a while longer.

@Bildo,- you may get your card up to 1040 and it runs rock stable at the moment, but damage is occuring whether you realize it or not. At some point in time, and it will be different from card to card, the card will fail while running too high a clock.

As a disclaimer, I'm not an engineer and I don't know all the specifics on this subject, but this is what I've gleened through the years concerning it. As such, this is just my best understanding and might contain a mystake somewhere.
 

Slobogob

Distinguished
Aug 10, 2006
1,431
0
19,280
The difference between 950 and 1000 is roughly 5%. If that causes instability, that means the card is already running maxed. Now if an overclock of 5% causes either severe eletrco migration or interference within the memory bus, then the problem is already present at 950 Mhz.
I am certain that an overclocked card will not last as long as a non overclocked, but a 5% overclock will not reduce the life span below a few months/weeks. If that was the case, the card wouldn't last a year even without the overclock.
The card will need a higher voltage over time to run at it's set speed. What could happen is, that the card is limited (by hardware or bios) to a certain voltage range. That would cause older cards to simply refuse to run once they are overclocked to a speed they managed when they were all new and shiny. If that would be the case, we will see a lot of people complaining about defective cards in a year. It makes me think about what nvidia did with the 9600 series cards.
It's not the 5% overclock that ruins cards over night though.
 

sailer

Splendid
Slobogob- You make some good points. I suspect that there is already elecromigration occuring at a 950 Mhz clock, but it is either deemed as acceptable or the company is gambling that the majority of people will change their video cards before it becomes a problem, or both. The problem with the 5% is that the card may already have been clocked by the factory to its edge of reliability, and the increased 5% may put it over the edge. Can't say for sure.

The higher voltage required over time and its effect is similar to what Anandtech talked about in a CPU article a few days ago:

http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3251&p=6

"Degradation-the process by which a CPU looses the ability to maintain an equivalent overclock, often sustainable through the use of increased core voltage levels-is usually regarded as a form of ongoing failure."

This article on CPU overclocking applies to GPU overclocking as well. BIOS limitations might well be considered a safegard to prevent us from killing the GPU or CPU even faster than we would otherwise.

You are correct that 5% overclock sould not kill a card overnight, not unless its already been stressed to its maximum safe speed. But it will hasten the death of the card. Whether that means the death will come within one year instead of five eyars, or within three years instead of five is anyone's guess, and perhaps a bit of luck.
 

sailer

Splendid
True, but lets say that I overclock my EVGA 7800 GTX and fry it. Does EVGA have any 7800 GTXs to exchange for the fried one?

As an example, I ran into a similar situation with Corsair once. I had bought some DDR 3500 LL ram and it failed. Corsair had stopped making DDR3500 ram, so I they ended up sending me some XMS DDR 3200 ram. If I wanted the 3200, I would have bought it in the first place and saved money. Thus the lifetime warrenty was there, but it didn't do me that much good.
 

bildo123

Distinguished
Feb 6, 2007
1,599
0
19,810
Ok, this does sound very convincing but how does BFG and other companies get away with selling there 88GT's with memory speeds of 1975 and I even saw one running 2000Mhz out the door? It doesn't seem like they retraced any of the wires etc...This is what confuses me...I actually have mine at 1975 and have had it there ever since JohnnyGuru released that post...BTW that 88GT that runs at 2Ghz memory out of the box is the EVGA SSC..And it from pictures it looks like a vanilla GT with just a different sticker on it...
 

LAN_deRf_HA

Distinguished
Nov 24, 2006
492
0
18,780
So what about gainward or palit's 8800 GTS? The one clocked at 730/1825/2100. Is there a big dif between the 512 and 1 GB set up or have they found that this isn't really an issue? I always got the impression that the new GTS and GT just had lower quality memory in general.
 

Hatman

Distinguished
Aug 8, 2004
2,024
0
19,780
Yep it IS true. Ive read about peoples cards being broken.


It cannot handle the 2ghz+ for long periods, which is why cards like zotac amp! use higher quality memory modules and not many manufacturers go to 2ghz memory.

Its only for 8800GT and only for 2ghz+ for extended periods of time. And the reason why my wc'd 8800GT is at 770core 1770shader and just 1950memory.

High memory clock for benches ftw.
 

sailer

Splendid


For both you and bildo, I'll offer a hopefully simple explaination of the how and why of the high factory oveclocked cards.

First come from when I worked as a mechanic while going through med school. Sears came out with a battery that was gaurenteed for life. Following that, some other companies, icluding the one I worked for, came out with lifetime batteries. I questioned a factory rep about it, as NO battery can last that long. His said that in the terms of purchase, the warrenty ended if the car (and its battery) was sold. Back then, most people sold their car within three years, thus voiding the warrenty. There was nothing really great about the battery. It was a standard five year battery with a higher price and a different name. The company gambled that most people would sell their cars before the battery died, and for those that didn't, the batteries sold for enough extra dollars as to more than make up for any losses.

Some some video card companies offer very high clocks on their cards. First of all, they charge more for these cards so they can recover the costs related to any card failures within whatever warrenty period is offered. Second, and this covers the lifetime warrenties as well, when dealing with high performance cards, they can figure that people will replace the cards with the next high performance card before the first one fails. How often do you buy your next video card? Do you really own and use it so long that if it fails in three or four years, you will be trying to collect on a warrenty? Or indeed, will you have replaced it before it gets that chance to fail? That's a gamble that the companies are taking. Their accountants have probably figured out the odds and told them how much they have to sell the overclocked card for so that they end up with a profit. The companies aren't stupid. They can predict a failure rate, and the cost to cover it. If by chance a card is especially bad, they just cease production and write it off as a bad gamble.

I can't say anything toward the actual physics of these cards, as I'm not an engineer, but I do understand that basic concepts. I will trust that if jonnyguru says there are troubles, he probably knows what he's talking about. I do think does make sense is that the finer the wire is, the more sensitive it will be to heat, electromigration, and such.
 

Evilonigiri

Splendid
Jun 8, 2007
4,381
0
22,780

I spotted the critical mistake... :pt1cable:

Sorry, I just couldn't let that one pass. :kaola:

Interesting analogy, just wanted to correct that the companies usually do statistics on how many people keeps the products for so many years and factors that in when determining the warranty.
 

randomizer

Champion
Moderator
I think it works kinda like this. Electromigration is where the electrons quantum "tunnel" outside of the wire. The problem with electrons is that they are so small that they don't follow normal laws of physics. You can never say "there is an electron here or there" since it can never be exactly pin-pointed. The probability of the location of an electron at any given point in time is spread across a range of space in a normal distribution. So in a typical wire, you would expect that you would find most electrons towards the middle of the wire and less as you move towards the edges. There is always a chance that the electrons are outside of the wires however. The thinner the wire, the higher the probability that electrons are outside of the wire. It becomes really bad when you have wires that are only a few atoms thick and you run into alot of problems as there are too many electrons "leaking" out of the wires.
 

sailer

Splendid


Alas, you have proved my imperfection. :ouch:

I do agree completely that the companies usually do keep statistics covering the average life of a card when in the hands of a purchaser and that many things go into the process of determining the performance and price of a given card, whether the statistics are studied by accountants or some other group, they will inevitably know more than I when marketing their product.

Randomizer did a pretty good job at explaining the basis of electromigration. You can add into it that heat plays its part, for as the wire becomes hotter, the electrons tend to "leak" at ever greater amounts. Also, as he pointed out, the thinner the wire, the greater the "leakage". In addition, the leakage from one wire will interfere with the wire next to it, and vice versa. It may well be possible that a point will come when higher overclocks cannot be achieved, as the leakage will be excessive. In that case, the chip companies will have to concentrate at having more work done during each clock cycle, or add more cores to do the work, or maybe both. Don't know the answer to that for sure, only speculation.
 

yipsl

Distinguished
Jul 8, 2006
1,666
0
19,780


What does RivaTuner show for your memory? My MSI factory overclocked 3870x2 is sold as 850/1800, but both CCC and RivaTuner show the memory as 901, and I read it's "effectively 1800" (probably because data's coming in as it's going out -- there has to be a circle in Dante's Inferno for tech company marketing departments).

I briefly clocked mine higher than the already good clocks, to 860/931 and it was stable in everything except for the game I'm playing the most right now, so I set it back. I don't plan on buying a different card in a year, just adding a 4850 clocked at 850 in CrossfireX when I upgrade to a new board and quad core.

Your signature's great, which is why I stick to computers, virtual sword and sorcery and imported Scottish ales. :pt1cable:
 

bildo123

Distinguished
Feb 6, 2007
1,599
0
19,810


Interesting as it is, depending on my memory speed set in Rivatuner, the actual speed read by the built in logging chart can go from one end to the other E.G. If I set my memory to 975 the chart will read it as 982, but if I set it to 973 It will read it as 970, kind of wierd but I kicked it way back down to 950 to be safe...Not only that, I bought a BFG 8800GT OC, and it wasn't really the OC version, GPU-Z/Rivatuner read it as 600/1500/900 before I even touched it for OC, so I'm not sure what happened there. Could that be considered false advertising :heink: ...I could use the money to pay off school... :kaola:
 

Hatman

Distinguished
Aug 8, 2004
2,024
0
19,780
There are levels, say for example, a custom oc of 555mhz core could be set to 510mhz because clocks go up in stages.

so, just for example

500 - 510 = 510
511 - 520 = 515
521 - 530 = 528

Just for example, search web for exact specs as idk.