M17X R4 - 980m - Throttling Research and Workaround

A couple of weeks ago, I swapped out my 680m for a 980m. As a piece of hardware, the thing is impressive to say the least. In terms of driver compatibility however (in the M17X R4, anyway), it leaves something to be desired.

At this stage, I'd call the 980m's compatibility with the R4 somewhat limited by driver support.

Regardless, as I have been testing a lot recently, I thought I should share my findings:

Relevant platform specifics: M17X R4, 3940xm @ 43/41/41/41, 32GB ram, 980m

OS Tested: Windows 8.1 pro

Bios Tested: Premamod bios (84.04.22.00.10 - card from @woodzstack)

Testing: Firestrike 1.1, Fallout4 (standing in front of a burning object on Ultra, 2k settings)

Installation Process: DDU, clean install (though I experimented with dirty installs outside these without results that were particularly different)

GPU Temperatures: 50 @ idle, 65 @ full load

Graphics Mode: PEG/Dedicated

Drivers Tested:

353.00 Peg Mod by J95: The best tested. Used without mods, the card holds on P0 until 99% load where it toggles to P1 and holds. Using nvidiaInspector, the P1 clocks can be changed to match the P0 clocks, making this more of a blip than a hard throttle. Included game profiles txt extends its usefulness beyond its original datestamp. Not compatible with VR headsets (deal-breaker for me, because I'm developing games for the Oculus Rift).

353.06 Peg Mod by J95: Comparable to 353.00

359.06 INF Mod by J95: Stable on installation with stable overclocking but goes back to throttling after a power cycle. After power cycle: Used without mods, the card holds P0 until ~90-99% load where it drops to what reads as P1 but actually has a clockspeed of ~324mhz. Can sometimes be tricked back into a pristine state by unreliable means (series of restarts, pin-mod cables, etc). Pristine state usually holds through reboots, but not through power cycles. Powerstate matching from above has no effect. Various additional settings (adaptive/high performance, batteryboost off, unnecessary services off, trying to reset the card by loading new user profiles) have no effect.

Benchmarking 359.06:
No throttle state (right after installation, 135/200 oc): http://www.3dmark.com/3dm/10227403
Power draw during benchmark (stock gpu clock): 183w-190w during the demo scene, max 214w on the combined test.
Power draw during benchmark (oc gpu clock):186w-196w, max 216w on the combined test
Throttle state: Radical throttling with the clock jumping from 1126, 1038 to 324. Not a lot in between: http://www.3dmark.com/3dm/10320712
Power draw during benchmark (stock): 182w-190w, falling to 112w during throttle periods during the demo scene. Similar data for the combined test.

For interest: Power draw with an underclocked processor @1.2ghz: around 163w. Doesn't help the gpu at all.

- Strangely, if you start the M17X R4 with a pin-modded cable, the CPU drops to its base clock (1.2ghz) but the GPU runs at full burn at what it calls P1 but actually clocks like P0. Benchmarks reliably but like one would expect with so little support from the CPU.

359.12 INF Mod by J95: Comparable to 359.06

358.50 by Eurocom: Comparable to 359.06.


Graphics Mode: Optimus

359.06, 359.12 INF Mod by J95: Comparable experience to PEG equivalents, complete with radical throttling.


That all said, if anyone's found a way to get this thing to run reliably under heavy load without turning into an expensive paper weight (aside from running drivers from the 347-353 era) please let me know.

 

Glitched Start - These are the peculiar steps I followed to get into a glitched/corrected state (where there was no throttling on the 359.06 driver without any other modifications). Note that I've not attempted to reproduce this and I have no idea whether this is stressing the components.
Mess around with this at your own risk.

Appendix:
Pinmod: Where you only partially seat the power-cable such that the control pin isn't making contact (or pull the pin out, or use an incompatible PSU). The computer (via osd) will report that the adapter is not supported and the computer will enter an unusual powerstate where the CPU drops down to its base clock and the GPU reports P1 state as a maximum.
Test: Firestrike
Power Cycle: Unplug the adapter and battery, hold power button for 30s, reconnect and power up.

1. Power Cycle for a clean start
   
2. Start windows
3. Pinmod - observe CPU drop to 1200mhz
4. Test - observe bad, fluctuating performance
5. Restart with pinmod still in place
6. Test - observe locked P1 state that performs like P0 w/ 1200mhz CPU
7. Restart
8. Unplug
9. Restart
10. Start into windows, Plug in
11. Restart
12. Pinmod
13. Restart
14. Unplug, Plug in
15. Test - observe good test with full marks
16. Restart
17. Checkerboard windows - graphical error where everything is rendering with a black pattern of squares overtop
18. Restart
19. Test - observe good test with full marks
20. Be confused. Sit down for a bit. Get a coffee. Question your life choices. Move to LA. Buy a dog. Name it "Mother." Find a passion for scuba diving. Get married. Have 1.5 children. Give the 0.5 to someone else with 1.5 children. Write a play about intelligent animals.

 

vbios 84.04.22.00.10 I think we flashed your card with.

I have always found 1 out of 10 drivers would work flawlessly and not throttle. The ones that did this, yielded about 5-10 % less power, but considering the 980M on the M17X-R4 is a godsend to begin with anything over 9000-GPU (not totally score, total was like 8500 or something..) score on Firestrike is good.

I would note, that you need to have more power. I solved this by using 3.9Ghz all 4 clocks, DDRL RAM at 1.5v only, and using low power SSD's, no CDROM, no lights/LED from Alienware CC, not using the onboard speakers- used 5.1 external sound, offloaded all the USB to a very reliable USB powered bar... I found just using a few more usb's or cycling the lights and speakers was enough to throw it over the edge and make the cpu throttle, and when the cpu throttles so does the gpu mostly too.

I also used a decent laptop cooler to keep the laptop cool, so the fans wouldn't even need to kick in, and used very good / excellent thermal paste like TX-4 or Gelid Extreme.

I've had a few M17X-R4's and this kind of trick worked on all of them. I started using it when the 880M debuted, and had insane throttle issues. 240W is just not enough for 4 dimms of RAM 3-4 harddrives, a 125W GPU and 100W+ CPU overclock, plus all the other bells and whistles.

if you really calculate the power when under full load, you will see, it adds up to much more then the PSU's.

The trick for M17X-R4, is not to use SLV's or any of PREMA's overvolted vbios unless your willing to sacrifice the extra 20-30W from something else or somewhere else. Also. 9/10 times the laptop would throttle the GPU in benchmarks, but very very very rarely in any games. When it did so ingames, it was only when the game just came out, was only a day old and was always fixed with a game patch, not a driver for the GPU.

 

You're right - editing post

 

Also - might I add that your score on firestrike is ridiculously more then to be expected. Its very high, if you compare that to what NBR shows the 980M range being, you would see your 10% past the highest point on the bar, so surely I didn't give you a dud of a GPU, and lets face it, you absolutely went beyond normal use, and what can be expected of it. Using a X6 and ignoring the fact I have a i7-6700K as my cpu and DDR4 and a m.2 2280 951Pro SSD in that, I would still not be able to easily get that high of a score, never mind just the GPU score alone.

So - congrats, your 4-5 year old laptop outperforms the fastest and most powerful laptops today for a 1/3rd the cost.

You might want to consider creating youtube video's or twitching your skills, and sticking to unlocking such potential and showing people, because it can, if you stick to it, make you famous.

 

Heh heh. Thanks, Woodzstack. It's by no means a dud card - thanks again for the quick turn around on getting it to me. (also, tell me about your X6 in a pm sometime - when I started this adventure, I was debating either upgrading the M17X or switching to the X6 or the DLX7).

Twitch: God... can you imagine a twitch channel of me restarting my laptop over and over again? "All right folks... restarting again *plays three chords on a guitar* here we go... uninstalling the drivers *plays three more chords* restarting with signature enforcement disabled *plays a chord, screws up the next*" -- and so on, and so on. I'll stick to writing code for the time being... that would also be an awful twitch channel lol. "Here you can see that the player falls through this section of floor if they have the menu open and are taking damage from an acid effect... let's try to fix that *6 hours later* and now the player doesn't fall through the floor!"

I'll add some watt rating numbers to the OP and break out the benchmarks for throttle and no-throttle runs.

Throttling on games, my tests: When the card is in a throttling... state/reboot/whatever... gpu load above 90% throws the clocks down to 324mhz. Now, you have to be pushing the settings pretty high for that (my Fallout 4 test was at 2k resolution, msaa, on ultra settings with a bunch of smoke particles on screen). If the load is less demanding (like say at 1080p resolutions) the card holds P0 just fine.

Throttling in general and firestrike scores: Man, I'd take a solid 8000 if the clock didn't keep dropping to 324mhz and holding there for a few seconds at a time! It's the radical fluctuation that makes it hard to work with.

 

Interestingly, the power draw during the benchmark should be well within the limits of the 240w PSU (max 214w while overclocked - measured using a Kill A Watt). The only thing that separates the Throttle and No Throttle tests above is that the No Throttle test is right after a fresh driver install and restart, where the Throttle test was done after a powercycle/cold boot.

Also, interestingly, the 3940xm (according to throttlestop anyway) reads a package power of 33w during the Firestrike test (compared to 60w doing an 8-thread cpu benchmark). Underclocking it, freeing up another 20w, doesn't affect the gpu at all.

Makes me wonder what changed in the driver code between the earlier 347-353 drivers and the more recent drivers. Heh heh, also makes me wish I could get a look at the source code.

 

SLV7 vbios helps to push that bit of extra performance though:
http://www.3dmark.com/3dm/10321003?

Edit - added 3d mark 11 score too:
http://www.3dmark.com/3dm11/10817423

I did listen to your advise in the end and got a proper OEM 3940xm instead of one of those ES chips.

 

Also - Power Management Mode (NV control panel). When the GPU is in its throttle prone state, changing this field to prefer maximum performance actually helps (but not in the way I'd expect).
In doing a test on the Rift (which puts a huge load on the GPU because of the stereo rendering), my test scene was pushing the card to throttle, where the load ran from 77%-98%. Same scene, on with prefer maximum performance set hovered around 77% load - I suppose preventing the load spikes by keeping to the higher power states? It almost seems like the card's instruction set has a condition that reads like "90% load? TOO MUCH". J95's thread at TI recommends this setting (but doesn't talk about what relationship it has to the throttling issue beyond making it better in combination with the peg mod drivers).

Note also that this setting prevents your card from hitting idle clocks / p-states... so if you're unplugged... consider switching it back to adaptive.

 

Hey, @mariussx - which vbios and driver combination are you using, for reference?

 

Hi, Schurke

It's 359.00 with J95 modded inf file and slv7 vbios (84.04.22.00.20) in a dedicated mode. I am trying dedicated only for now to see if my screen is still getting EDID corruption, since it seemed easier to get it than using Optimus mode.

 

And are you experiencing throttling under heavy load or is that combination working on your rig? If no throttling, are you using any special steps to get there (like J95's state matching technique etc)? (Sorry for the storm of questions - piqued my curiosity now)

 

Nothing whatsoever really Just DDU all the graphics drivers and installed the modded driver as usual. No matching P states or anything at all, I just set the 3D mark to maximum performance in nvidia control panel - my GPU never throttles (got it from German gaming laptop re-seller); if it does at any point - I just repeat DDU and driver reinstall. When overclocking the GPU I am trying to use the lowest voltage possible for the benchmark not to crash. And of course I only use those clocks for the benchmark to see if the latest drivers are any better than previously released. Clocks in games are either stock or slightly higher if I need some extra performance.

Temps and no throttle clocks below:

 

That makes sense - it sounds as though you maintain that fresh install state for as long as you can, then just re-install the driver if you lose it. Not the worst thing. I'll give that a try for a bit and see how long it holds between re-installs (given that I use my M17X as a laptop - in and out of my bag, on and off the wall charger).

 

Edit, just tried 3d mark 11:
http://www.3dmark.com/3dm11/10817423

Well, it stays in no throttle state for months... The only time GPU start to throttle is after new driver install sometimes, when I have to re-DDU again (sometimes even twice - it would throttle again, so DDU and re-install again etc) and then it's back to no throttle. I forgot to mention my PSU is 330W (upgrade over original 240W, which came with 680m at the time). If you are using 240W PSU, I believe that is your main issue with throttling... R4 does not need any mods for the PSU, genuine DELL PSU (Dell part number XM3C3) works perfectly fine. I my case, the laptop stays plugged in most of the time as well, I almost never use it on the battery.

 

I'm using driver 352.84 on windows 10, 120Hz display, and a 240W PSU. CPU is OC to 4.3 GHz. I think a 330W PSU is a must.
After installing the XM processor, I started getting driver crashes on SVL7's vbios (1.00v). Now testing with Prema's vbios (1.062v).

 

So, you and @Solo wing support my max power draw theory, BUT - the OP still has pretty awesome performance regardless.

 

@woodzstack , @Solo wing , @mariussx - Do you guys have access to power draw measurement tools (like a KillAWatt or similar)? Tests at my place showed most of the Firestrike benchmarks not even cresting 200w (and underclocking the CPU I was able to get down to about 165w draw and still was throttling after leaving the "fresh install" state). I'm wondering now if my readings/tool are off and if any of you can verify... Because if the draw is sub 200w, a 330w PSU shouldn't be much more than a placebo... right?

 

@Mr. Fox did a power draw test of a single 980M, it never went below 270 during the test:

 

so 10-15% OVER the 240W PSU.. minimum. interesting. Well, sure puts things into perspective.

 

Unfortunately not, I don't have the tool to measure live power draw. I am pretty sure even when I did not have xm CPU, but 980m alone the 3d mark scores were worse until I got the 330W PSU. In fact, I should still have 240W PSU somewhere - when I find it, will try to test with both PSUs to see what the difference is. It will have to be in the evening or tomorrow.

 

Please do - thanks Mariussx.

After that, I suppose I'll try to strip the machine down and see if it runs differently bare-metal. It seems strange that with the CPU radically downclocked, the issue still persists.

That said... the PSU is now more that 3 years old. Output/efficiency does degrade over time. What the KillAWatt is reading is not necessarily representative of how much energy is actually being converted, just what's being drawn from the wall.
(That said, Fox's numbers there are crazy high in that video... as though he was also running the mad, benchmarking CPU overclock he was known for).

Given that I'm experiencing radical, clock-hitting-324mhz-and-sticking-for-a-few-seconds-throttling, it might be advisable to check into regardless. From what I've read, this is not the normal sort of throttling people experience with this hardware configuration... (correct me if I'm wrong here).

Also, @mariussx - you're running premamod vbios 1.1.1, right? That one allows undervolting, right?

 

Oh, nearly forgot. I did another install of earlier drivers this morning to see if I could notice anything unusual going on. In 350.12 and 353.06, what actually happens during the benchmark is the GPU starts in P0 then drops into P1 as soon as the utilization hits 99%. After that it mostly stays in what nvidia inspector reports as P1 - HOWEVER - it remains clocked at the stock frequencies and voltage (rather than say, the overclocked P0 frequency). The 359.x drivers will go P0 until 99% then drop to what it reports as P1 but clocks like P8, mostly switching between (for a few seconds at a time) 1127mhz and 324mhz with little in between. These behaviours are consistent across restarts and cold boots.

 

No, SVL7 vbios:
https://www.techinferno.com/index.p...vbios-mods-900m-series-overclocking-versions/

It allows both overvolting and undervolting:


Edit - as promised, tests with both PSUs.

Ok, running the tests on stock clocks (svl7 vbios stock is 1202 core and 1253 mem), no OC, no settings changed 330W PSU first:
http://www.3dmark.com/3dm/10340072?
55.7 max TDP, 64C max recorded temp

240W PSU run, all settings exactly the same - it seems on stock clocks 240W PSU gives enough power:
http://www.3dmark.com/3dm/10340244?
55.9 max TDP, 64C max recorded temp, still no throttling at all

time to check the OC now


Sure enough 240W with OC hits the power limits and 980m throttles like hell:
http://www.3dmark.com/3dm/10340422?
nothing changed here except for the OC (power/temp target set at 197%), 89.6 max TDP


And the last one for a good measure - 330W PSU re-connected, nothing else changed:
http://www.3dmark.com/3dm/10340591?
91.2 max TDP, 71 max recorded temp (card downclocks to stock between the tests as per screenshot below, but no throttle in the actual test)


Almost forgot, the last thing to check is your card's ASIC quality in GPU-Z (although I heard it should not be a problem with a card from Woodstackz in your case)

 

Cool - I'll give other vbios's a shot after you get back regarding the 240w.
-- I should have read back - you actually did all ready tell me you were using the SVL7 vbios. Sorry about that!

 

@mariussx Wow... Yeah... there you have it. Damn.
This would support the theory that the configuration pretty much maxes out a 240w and if there's any age/wear on the PSU... it's not going to deliver. I fired up GPUz and did a test with it open. The PerfCap is almost all green (pwr). Heh heh, I think you got me, Mariussx! I'll order a 330w pronto and I'll eat my "the readings don't support it" doubts... if it works. Thanks for the testing there - really appreciated. This "sometimes it works and sometimes it doesn't and there's no real rhyme or reason why" stuff makes me crazy.

Oh ****. That could also be why the pinmod garbage from my testing works... people were using that to crush power out of the old M15x PSU... Pure speculation at this point. I'll stop talking until I get a 330w.

Also, while I have you are you using A13 or A11 unlocked? I ask so that I can revamp the OP later with other details, in case anyone else with the same config comes across it.

Also, thanks @woodzstack for contributing and adding the supporting anecdote about the 880m power throttle.

 

No worries At the moment A13 stock, never got around to flashing A11 unlocked BIOS.

 

You're not missing much. Most of the options do nothing on Ivy Bridge (you probably all ready know this). AFAIK all it lets you do is swap peg/optimus from bios.

 

You might not be seeing a high power draw if you are battling throttling problems. Trying to run 980M SLI in the M18xR2 drew crazy amount of power in the VERY RARE circumstance that it did not throttle on me. When it was throttling it wouldn't even pull more than 330W when overclocked. I finally gave up and bought a different system to use the GPUs in because the throttling was insufferable.

 

I read about your fight with that. I'm sure it was a massive let down after working so hard on your R2. That sort of thing leaves me feeling gutted.

Hopefully in x days I'll be asking to get this thread re-titled to something like M17x R4 w/ 980m - get a 330w PSU (rather than M17x R4 w/ 980m - the compatibility crap shoot).

 

I hope you can get this working soon. M17x R4 is one nice machine. Ivy bridge (unlike my nehelem) still really current in terms of performance.

Shortly I will be installing a 980M in my M15x. Lets compare benchmarks. I warn you I use dual PSUs

Should be fun to compare old vs ancient alienware from the golden days of the brand.

Sent from my SM-A500FU using Tapatalk

 

I really hope 330W PSU fixes all your throttling problems. Just a quick update on the 3D 120Hz screen install (quite easy by the way) - got it installed today and I had to re-install nvidia drivers. I don't know what it is with these drivers, but I had to through DDU - install the 359.00 driver - set max performance in nvidia cp and 3d mark run procedure for like 4 or 5 times. The card throttled like hell on the first installs and although I did not change anything whatsoever, just repeated the exact same procedure, I got it working finally. It was really frustrating at first, but finally it's back to normal - no more throttling. There might be a way to fix it with a p state matching, but in all honesty I did not want to go through it since I knew it was working fine before and it can be achieved by simply re-installing the drivers. The added bonus is the card never throttles in any game or benchmark afterwards, which kept me going.

3d mark 11 got 15.7k points in the gpu score
http://www.3dmark.com/3dm11/10837854

 

@King of Interns - you should talk to @mariussx for benchmarks - he seems to have things pretty much sorted out. The M15x still hits above its weight... and is weirdly up-gradable. I've heard that it can field a 970m or 980m (while even the M17x R3 can't). How does that even... work? I have a largely stock M15x at my place still that might inherit my 680m (if all of this gets sorted). Is there a thread about that particular upgrade?

@mariussx - That's some pretty spooky behaviour... I'm waiting on the mail (probably another week) for the 330w. When it comes I'll post some notes and GPUz sensor screengrabs. We'll see what happens... That said, your benchmark is VERY impressive - congratulations! Now, when your card does start throttling, is it bottoming out to the 300-400mhz range or is it just jumping down to the base full clock (whatever it is... 1000 and change) or some other arbitrary range? Anecdotally, the p-state matching didn't do anything for me when my card was bottoming out on 359, but did work as described on 353.

 

Sorry, I honestly did not even check what was the throttling like- as soon as I saw the FPS dip in 3D mark way below what they should be, I quit the testing and wiped the drivers. Then installed them again. After about 4-5 times, they fixed themselves - there was no more throttling whatsoever. I ran firestrike and 3d mark 11 + some games and restarted/cold booted/woke the laptop up from sleep multiple times to confirm the drivers are installed ok this time - everything seems fine now. Card never throttles anymore - it was still quite frustrating to have to reinstall the drivers so many times. I am happy it's sorted now and I know there will be no more throttling until the next driver re-install.

 

Bought a 330W PSU from Dell. I can officially say that the 240W PSU is not enough to power an Oc'd 980M (1202/3000) + 4.1 GHz 3940XM. 1100 points increase due to the PSU upgrade is great!
These are latest results:


I'm hoping to push into the 10k for Fire strike.

 

Hi, @Solo wing - I'm still waiting for mine but that is great to hear - Congratulations on the wicked score!
Can I ask which vbios you're using? Also, has the 330w allowed you to move on from the 353 peg mod driver? (I think I saw you mention you were stuck on 353 - possibly on the Tech Inferno thread).

 

Cheers mate,
I'm using SVL7 vbios:
https://www.techinferno.com/index.p...vbios-mods-900m-series-overclocking-versions/
I like how the core is OC'd at 1202 @ 1.000v so the card runs super cool. I did not try another driver yet, 352.84 + J95 profiles are working great.
Driver 352.84 is for Win10, 353 is for Win 8.

 

I'm also considering the 330W . That's a really nice score

 

So... My 330w PSU arrived today. Before I even start, I'm going to frame this by saying that I just had a really... weird... breakthrough?

Let me explain as an anecdote:

Last week, my card just started working correctly for no discernible reason. By correctly I mean that the clock speed stopped flopping whenever I hit high load. I noticed when we were doing some demos on the Rift - usually I give testers a disclaimer before the run, saying that my computer is a bit botchy blah blah blah frame rate may not be ideal. That day, it ran through all of the demos without a hitch (un-optimized, photo-realistic room, two eyes, 75fps sustained). That night, I checked out Fallout 4. No clock/framerate drops even at 99% load (Ultra is pretty demanding). Firestrike run - piles of PWR performance caps, but no clock/framerate drop. Looked like this:



Remembering what @mariussx said about his good state (how he would get to it using repeated applications of DDU - but that it would hold after when he used his machine primarily docked) I changed my behaviour a bit and started using sleep and restart as my power cycles, avoiding shutdown. Card ran perfectly for two weeks like that.

SO... what happened? Scratched my head on that one for a bit, then remembered that prior to changing my habits, my machine hibernated one night (left it in my bag, didn't take it out until the next day). That was the only thing that changed.

With my new PSU in hand, I decided to confirm my findings. I performed a cold boot (battery out, discharge) on the new 330w PSU.
1. Firstrike test: GARBAGE.



Okay fine. Disappointing after waiting so long on the PSU but not entirely unexpected.

2. Hibernate.
3. Wait 5 seconds.
4. Hit the power button.
5. Machine wakes up fine. 3D mark and GPUz are still open.
6. Firestrike test: Good.



7. Cold boot - repeat whole process. Test: **** show. Hibernate. Test: Good.
8. Repeat again, this time using the 240w. Test: **** show. Hibernate. Test: Good.
9. Either PSU. Restart after good run. Test: Still Good.

(both good runs: max volts 1.03, peak draw at outlet 230w)

So... I'm at a loss as to why this works on my machine. I'm still using the prema vbios Woodzstack shipped. Still using 359.06. CPU still at 43/41/41/41. Still windows 8.1. Literally the only difference is... Glitching the card into behaving by exploiting hibernate, apparently?

I did a limited OC run (and then afterwards a power-state-matched OC Run) of firestrike (limited because the vbios I'm using is limited to a 135mhz oc). Clocks held good-ish. No peculiar behavior. Power-state matching kept the card from dropping to its base clock (1035) and kept it between 1170-1260 for the test. (note - visibly smoother but doesn't seemingly do anything for the score above a normal oc).

I am... going to flash prema mod 1.1.1 to see if it exhibits the same behaviour on my machine. I'll post again later.

In the mean time, can someone else with a similar config confirm my findings at all? @mariussx , I wouldn't ask you to muddle your machine on purpose, but if you happen to drop your state... maybe give it a go. @Lord Fisher - I saw your thread... you might want to try this and see if it works for you. If you do... please post here.

Oh - if anyone tries to reproduce this - the only other setting I can think of that may or may not matter is that I have "Prefer max performance" set in the control panel

 

Premamod v1.1.1 addendum.

0. Backed up the old vbios
1. Flashed the card.
2. Cold boot.
3. Firestrike test: garbage



4. Hibernate.
5. Wait 5 seconds
6. Resume
7. Firestrike test: perfect
8. Apply sloppy OC (200/200 - guestimate based on baseline .03v bump on this vbios)
9. Firestrike test: perfect

...like... eerily perfect. It held the clocks absolutely, without the application of power-state matching.


http://www.3dmark.com/3dm/10857036?

10. Restart. Test. Still perfect.

Notes:
Card runs about 10 degrees hotter using this vbios
Peak power draw at outlet: 247w

 

Did you try SVL7 vbios yet? I have been using them for just over a year now and could not be happier. 76C seems a bit high (if ambient temp is around 22C) for a quick Fire strike test.

 

I'll give it a shot to complete the exercise. I did a fallout 4 test after on the 1.1.1 vbios (or rather... instead of going to bed I played fallout 4 for a couple of hours). Temperatures were about the same (76-78) with loads at 90-99 - which you're right - is a bit high for an all-the-time overclock. Also, it would be nice to not have to marry the 330w (at 1.06v you're outstripping the 240w at peak).

I'm now also curious if the same hibernate exploit will work on later drivers (though people have had a hell of a time with 36x across the board so I might hold off on that until I have a reason).

 

SVL7 vbios addendum

1. Flashed the card
2. Cold boot
3. Firestrike test: garbage



4. Hibernate
5. Wait 5 seconds
6. Resume
7. Firestrike test: perfect



8. Firestrike test 240w: perfect



This vbios is probably more practical for walking around - at its stock, it doesn't appear to out-strip the 240w (or rather, with my system components and cpu oc it doesn't).

That all said - the hibernate exploit... it sounds more like a driver bug than anything else. Also it sounds stupid as crap and like the sort of hocus pocus solution I'd normally just ignore. Until anyone else reproduces the test, all I can really say is "works for me". However, I am going to tag a few people to get some eyes on and opinions (people who have weighed-in previously or fought with this platform previously) - @woodzstack @j95 @Prema @Mr. Fox @Solo wing

 

Hmm that's quite weird. I naturally don't advise touching hibernation with a ten foot pole.

 

I have been using 352.84 + Windows 10 since June/2015, I truly gave up on all the other drivers. +1 @woodzstack I want to keep away as far as possible from Hibernation/Fast-startup.
While gaming, the card performs great, (1202/2910) no need to P-state match. Firestrike on the other hand causes the core clock to switch to P1 (1126/2910).

 

Are you using our custom vbios or someone elses' ?

 

Using johnksss & svl7 vbios, the card runs great and super cool @ 1.000v. Prema's v2 caused lock-ups and ran very hot.

 

SVL7's is higher voltage when it is used and in turbo, whereas Prema's I think its a more steady voltage at all clock speeds, from what I understand. Someone correct me if I am wrong.

 

From what I saw, the SVL7 vbios runs at a steady 1.000v. The one you're shipping changes with the current frequency (between 0.975v and 1.030v). The prema 1.1.1 vbios runs at a steady 1.062v (which is interesting because that voltage paired with an 3940xm pushes the power draw to about 247w - might explain why some had so much trouble with it).

I'm actually now wondering about the voltage core on initialization. Here's why:

On the 2 prema vbios...es that I tried, the core voltage is set north of 1.000v (1.03 and 1.06 respectively). On drivers outside of 350/2/3, the card initialized strange after shutdown and throttles massively under high load (but can be cheated using hibernate, seemingly). Now, the SVL7 vbios is set to 1.000v - cold boot, the card runs like poop and can be cheated in a similar way. However, on this vbios, the cheated state survives a shutdown (not a cold boot, but coming back up from shutdown, the card still runs fine).

It makes me wonder if an unlocked vbios that starts at a lower voltage and clock (say, 1035 @ 0.975v) would behave more consistently on the newer drivers. Does anyone know if a low voltage vbios variant was ever made?

 

So it seems the GPU today was in the mood to allow for core OCing Finally in the 10k range!(1350/2910)