GTX 970M Voltage/Overclocking Problem

Hi to all,

again i have a problem with my beloved p370sm-a, especially the graphic cards. When my 980m died i bought 2 970m's from china as the prices here in europe were to high for me, i bought each card for round about 250€.
However, these cards work good(@ stock clock), after 3 month no real problem.

But when it comes to overclocking one of the card fails - they both have extremly low ASIC values-(62,3%, 55,8%).
The card with the lower value easily clocks to 1170 mhz@ 1.043V core. the other one with higher asic fails with [email protected] value is 1.025V.

I tried to modify my bios and with higher voltage, then the card is more stable, but not efficient, which means lower 3D Performance.
@Prema mentioned an Asic Interpreter for the voltages in a post about an 970m custom bios - could this function been broken on my card?

Is there anything i can do about it? can someone explain how the Asic Interpreter works?

Thank you.

 

Ignore the ASIC value, it just appears you got one card better than the other.

 

yes, that's what i initially thought, but in this thread:

https://www.techinferno.com/index.php?/forums/topic/2442-biosvbios-mods-prema-mod™-stock/&page=109

Prema mentioned this Asic interpreter which is directly messing with the voltage and I thought maybe there is a way to find out how this works.
By the way, this vbios, among other Prema vbios versions ist still available for download, but I don't want to bother Prema with it..

 

Sorry but don't.

Seriously don't overclock these cheap chinese cards, you will only kill them extremely fast. There is a reason as to why those cards are so cheap. My old GTX 970M Overclocked easily to 1.455mhz on +75mv to give you an idea how bad the cards are you have.,

 

Hi Danishblunt - yeah, I know these cards are rejected probably from dell (pcb-layout) because of minor irregularities and low asics during manufacturing process and are mainly for chinese or south american markets.
but they are not defective - and with good cooling overclockable as any other card. I love to tinker with stuff like that - my 980m died prematurely - an original clevo card!
nothing justifies the price the reseller told me - 850€ for a replacement card!

my other 970 with +70mv reaches 1342mhz stable - which is not too shabby for half the price.

But thank you for your interest.

 

You'd be suprised. the cards will likely die in matter of weeks if you keep them OCed and stressed out. Also I buy MXM cards from Europe all the time and I never heard of 850Euro for a GTX 970M. Infact I bought my 970M back then for roughly 330Eur. The chinese cards are not made to withstand high clockspeeds or voltage, it has nothing to do with silicon or defective.

You have been warned, if you suddenly have a black screen and fans going 100% speed while your lights are flashing, then you know 1 or 2 cards are dead and you know the reason.

 

The 850€ were the price a german well known clevo reseller told me on phone for 980m in january this year. look at the prices eurocom has for a 970m - ridiculous.

I know a thing a two about electronics and have overclocking experience of nearly 20 years now - why my clevo card (and many other 980m) died at the age of 1.5 years if it's quality what matters?

I never killed a card with overclocking - there are many factors to consider before pushing hardware to the limit.

 

Just to point it out there, I literally bought a GTX 980M 8GB card from Europe for 306 Euro a couple of days ago. It's coming tomorrow or the day after tomorrow so I really don't know where you got your weird sources from.

Here on ebay as well: https://www.ebay.com/itm/Genuine-Nv...345870?hash=item3d5c77950e:g:XHsAAOSw2n5ZiIJ-
New, not even close to the 850EUR you quoted.

Also why your card died, I wouldn't know, I don't know you as a person and I don't know how you treat your systems. I have thus far never killed a graphicscard so far except for the trash designed Fermi cards (they were designed to die), I even have a macbook Pro 2011 model with the notorious AMD GPU still working flawless never needed to get repaired in its entire lifetime, my GTX 980M which is overclocked on my MSI notebook works since 2015 and is working nonstop for 6-8hours a day for 3 years.

I have had way to many people who bought chinese maxwell cards and always when they tried to overclock them they died on them in a really short amount of time because the components on the cards are usually extremely cheaply made. Much like the Gecube cards when trying to raise the TDP just a little bit they also die shortly after.

In the end those are your cards, your money, I simply warned you. If they break because of the overclock and voltagte adjustments then so be it.

 

hi,

first: you don't seem to know much about electronic manufacturing processes,
second: 490 Dollars are 424 Euro, not 306, a clevo reseller is not some guy on ebay.. Try MySN, or Eurocom, Sager or someone else next time.
third: You have an opinion on this topic and i respect that, so don't offend me with such a tone.

My question had a technical nature - I want to know how things work.

 

I believe the voltage adjustment would be a higher\lower voltage at the same clock speed. Basically an offset.

 

1.) You don't seem to know much about it either, otherwise you'd be able to fix your broken 980M.
2.) I just told you that I bought a 980M 8GB card for 306 Euro, and then proceeded to show you another link showcsing that you can get another card for 424euro by me literally searching on ebay for 5 seconds. Why would I go buy from a reseller like MySN, Eurocom or Sager if I can buy if from a private reseller who stocked them up from Clevo direcly?
3.) There is literally nothing in my post that is offensive in any way shape or form.

Also the ASIC in GPU-Z is as reliable as windows experience index. Using silicon leakage binning fuse to determine the quality of a graphisccard is the same as using your heatsink temperature to determine your CPU core temps. Take the value with a grain of salt. The reason why your other GTX 970M despite better binning is worse is mainly due to inferior VRMs.

 

The 9xx series is not hugely sensitive to vrm quality. Its likely just silicon lottery variance.

 

man, for the sake of conversation..

every electronic component on this card has tolerances in production of up to 20 percent which also can change in the soldering process - thats why you always have a part of production being rejects.
you can't test every resistor, capacitor and so on for being in tolerance before the product is finished - thats why there is quallity control. most cards produced are in that tolerances, some are not. it's a natural process. Prema mentioned an Asics Interpreter which uses the programmed Asic score to change the voltage of the card - that was my interest. this guy from ebay is selling used stuff - look at his other offers.
yes, he offers a warranty - but what happens if you need to rma the card from finland? he stocked them from clevo directly? - Sure. And when you have paid delivery cost and import fees your card is much more expensive as it's listet - depends on your origin.
there is nothing like 'counterfeit' mxm card - wrong labled maybe, like the 1050's for 60 euros from china, but you need a factory with relative high standards to produce such a complex thing like a graphics card - no backyard soldering.

I wanted to buy a new card.

in electronics production you have to have minimum standards - otherwise your whole production is doomed. most expensive parts of the card is the gpu itself, the samsung ram, and the pcb.
they don't save money by using cheaper parts like the vrm's which cost the factory a few cents each, as they need thousands of them. they save money with using cheaper gpus and remaining stock of vram.
the cores of the 970m were also used for the oem version of the 960 desktop - the cheapest production of gm204 as these were never meant to be gaming cards.
but these 960 oem cards have much better cooling solution than a notebook - thats why you need the core voltage in a notebook as low as it can get because the cards have to be reliable at least for the warranty.
The Asic Interpreter is doing exactly that - a programmed value which keeps the card stable and reliable. Read Premas description in the techinferno post.
I think they used these cores for 970m's as there is still a market for the mobile cards, but not for 960 oem stuff. that's china.

look at this one - no sli connector, not necessary for a single card laptop, but i have never seen one before - thats china:
https://www.ebay.de/itm/Alienware-N...401700?hash=item1cad0d1ba4:g:kW4AAOSwtG9bDRVf

my 980m, yes i failed with the soldering of an integrated circuit - bad luck which always can happen, lead free solderings is new for me - but at least i tried and almost did it.

my attempt is not breaking an overclocking record with my cards, but to have a stable ~20% 24/7 reliable overclock. thats not impossible.
yes, my cards are inferior in some way, but not how you think - and i am glad to have 2 new cards for the price of one - even if they are not as overclockable as hell.

 

Wasn’t it lack of VRM’s that killed so many 980m ? I mean @Mr. Fox killed 3 of those graphics cards in a row. Many people added VRM’s to save them from the death if I remember correct.

 

@Papusan:

yes the lack of them killed the cards, not the quality. 980m's run too hot for three vrm's as can't deliver enough amps above 70 degrees for a long time and degrade very fast.

it's case sensitive how good the cooling works.

 

To save it from power shutdown in my case but they were running at 50% overclocks and probably drawing over twice the power of stock. I think that's reasonable that that might push the original design.

 

The 980M and 970M were very capable cores. My own 970M got to ~1340mhz before the VRM just couldn't keep up and I was only putting minimal extra voltage into it. Temps weren't even a problem (low 70s).

From what I can tell (my knowledge here is limited), the original 970M MXM cards used a 2+1 phase VRM, but each phase used 2 mosfets. That would put less load on each mosfet and spread the heat load out. The newer green 970M cards cut back to single mosfets which the 980M cards also use.
It would seem to me, if you could mod a 980M board to run dual mosfets (the pads are already present, not sure of traces) you could very well end up with a card that would out-run the desktop 980 rather easily. If you can keep the VRM cool that is.

 

Prety sure @Mobius 1 has already done this kind og mod to a GTX 980M

 

TFW when a BGA 980M has better VRMs than the MXM version:

 

haha twice as many mosfets. FeelsBadMan

 

Not my picture, but yeah that's the laptop (P650SG) I had previously. Ran the core at 1351MHz 1.162V and 6GHz mem 24/7 for 3 years before trading in for my current system. It beat a stock desktop GTX 970.

 

it was done by @Khenglish

 

thought it was you since u sold the card. My bad.

 

You got better memory overclocking with less chips and under normal loads the efficiency is better.

I believe I was the first to actually mod my cards to have them fully populated:



Also placement and cooling have a large impact on VRM performance. The VRMs in the BGA versions did not have heatsinks, just airflow from the nearby fan.

As usual there is some engineering nuance going on here, think about why things are put that way from an engineers perspective before jumping in.

 

Do you have old benchmarks on how high it scored? Would like to see what 6 mosfets can achieve.

 

It made little difference, you got a bit more core speed but lower memory speed due to more interference. What it did mean is that I could benchmark several times in a row with an over volt+50% overclock and not have the card shut down on me.

https://www.3dmark.com/fs/3691881

Was an example, notice the 1500Mhz memory clock, before the mod with a vmem mod (which helped keep memory clocks at 1500Mhz post mod):

https://www.3dmark.com/fs/3356931 1650Mhz mem clocks
https://www.3dmark.com/fs/3593223 1550Mhz mem clocks

Post mod:

https://www.3dmark.com/fs/3689942

So a few extra hundred points because of no shutdowns.

 

Actually on the P650SG they did have thermal pads touching a copper heat plate and vent on the bottom cover.

 

Not the SG:

http://forum.notebookreview.com/thr...l-impressions-and-review-by-htwingnut.767349/

 

Yes the SG. I had that laptop for over 3 years and took it apart numerous times. There's no picture of the bottom cover with the thermal pads and copper sink/air holes over the VRM area in that thread.

 

4th image:

 

Where is the picture of the bottom cover? SMH.

 

I thought you mean this heatsink design used for the 10xx cards:



But I see you meant:



Which yes is not totally bare, but still not on the same level as:

 

Bingo

 

So still the extra chips there kept it easier to spread that heat out over a less efficient heatsink. Good design said put them in, truly active cooled can make a big difference hence less put in, helps idle power numbers and has less EMI and lower cost.

 

Thanks for making me aware!
I have initiated a follow up to my original take-down request from over a year ago!

Those old Mods are no longer safe with the trash MOSFET cards they started selling later on with higher ASIC DIEs.
A higher ASIC card will use less default voltage and hence the MOSFETs it is shipping with are not tested under the same stock load as the low ASIC cards from older batches.

Then later on they reduced the amount of MOSFETs even further...

 

I modded 4 980m boards to have all core FET pads filled and never saw a memory clock loss. I did also fill in many of the empty supply voltage cap pads too though on the 3 I tested the most, so maybe I would have if I didn't do that.

In general the 9x0m boards clock memory rather poorly. With the memory at 1.5V, 980 memory chips, and 980 timings the cards would need to be kept cool for just 7.2 GHz to be stable (7.6 benchable with crazy artifacts), while a 980 would do 8 GHz. There has to be some sort of capacitive coupling or lack of termination issue with the 9x0m boards.

 

To be fair I had volt modded the memory chips so I was running them beyond what most people would reach for their testing.

MXM boards are more cramped so tracing is more challenging that desktop boards. My 780m cards would reach 1700-1750mhz iirc with volt modding them which is a fair bit higher.

 

Sorry that I'm a little bit off topic. I have a 970m clevo version that I installed Prema's vbios. On running furmark, it takes less than 3 mins to blackout. Can I contribute it to MOSFET failure?

If yes, what can I do? It still runs most of my games, fails from time to time.

 

Try replacing the thermal pads, checking their fit and backing off any overclock.

 

Later 970m cards were cheaped out with the power FET count cut in half, and they cannot safely run Prema's bios which has the power limit removed. You should revert to the stock bios or else your card will likely fail. The blackout is an emergency shutdown of the GPU due to power FET overheat.

 

You say cheaped out but at stock it would have reduced power consumption, at idle too.

 

Many thanks for both your comments.
I'm pretty confident that is an overheating problem. Is under-voltage an option? or better MOSFET cooling?
Before reading your replies, I raised the voltage to 1.1v to emphasize the problem. shut down time reduced to 1min.
Maybe I should keep it 0.95 or even 0.9?

 

You can't undervolt. Just revert the vbios flash to restore the power limit.

@Meaker@Sager

Yes 2 FETs will burn less power than 4 at idle for sure, bit I can't see how 2 would be more efficient than 4 at full load considering how close even a stock 970m with 2 FETs is to the FET current rating and how inefficient they will be running.

 

It's not that close at stock, it's just overclocking these chips ramps the power consumption.

If you want to push it you can just get the two chips soldered on. Digikey will stock them.

 

Where can I find stock vbios? is the following link good ?
https://www.techpowerup.com/vgabios/165701/clevo-gtx970m-6144-140905

 

Sorry if I interrupt,
I do have my 970M Clevo blue pcb 6gb with unlocked vbios from techinferno (because of the lower stock voltage compared to Premamod v2). It runs at 0.993v at stock, and it's stable at 1261mhz at 1.025v and 1300mhz with 1.05v.
I do not stress it, barely use it for light games like PES wich do not need OC to run. I just Oc it when I play racing games like project cars 2 and forza motorsports. It never passes 86C at full stress in these occasions, ASIC is 65.8%.
Do I have to worry about frying it in anyway?

 

After flashing stock vbios, running furmark seems ok. at first it runs on 0.875v 750mhz, due to power limitation I believe, then back to 0.985v 1150mhz.
Everything seems to be fine. Just curious, what is power limit of stock? 75w?

For oile's question, I have no idea. Intuitively, if you don't experience any problem then probably it is fine. The voltage is not a direct cause but power comsumption. That's what I learn from this ordeal.
Thanks again for your help

 

As khenglish said, then card has some severe MosFet issues, I had the GTX 970M myself and noticed these issues as well. Had to make a custom vBios that would find the happy medium between performance and voltage. The powerlimit should be around 75W, yes.

@ oile
Temps are a little to high, try to make it so it never passes 84c. Preferbly around 70s. Other than that it's fine, as long nothing crashes you're usually good to go.

 

It reaches those Temps only when the tdp gets crazy high in only very few games (as of now the only one that seems to push the card tdp to over 100w is project cars 2 when raining, I don't know why).
Otherwise it never reaches 75C,
I could be at 1172mhz with 0.962v and a 85w tdp normally, I should check now because my memory is not serving me right.

I remember once a time in wich I was testing extreme Oc with extreme tdp on project cars 2 the screen went black, it was the mosfets protection right? Or did I damage the card that time?

I do have these two cards at home. I am using the Clevo blue 6gb one. I guess it is one mosfet per phase, so how high can I go in tdp before damage?

 

It's just the VRM shutting down due to overheating, it should shut down and not die. My 980M modded for higher end overclocking: