SLI MSI 1070 MXM in P370EM... An adventure

Pic hoster changed to imgur and first few posts relinked.

Ultimately you may be right and this issue may not happen anywhere else ever (certainly I haven't had any reproducible crashes in the month since I put the cards in) but what I've done so far to reduce crashes hasn't changed the game in any way

GTAV is the first decent game test I've done, and even then 1070SLI is such overkill for 1080p60 the cards just aren't getting pushed hard at all

 

Swings and roundabouts.

vBIOS power limit mod worked, using the preset in @Coolane's tool the power slider is unlocked and it draws more power. It would hold at about 1260mhz in Kombustor at stock 115W, on a (very short) run with the power limit at 100% it looked like it was drawing 150W and getting around 1560mhz in a single GPU test which was stable for 20seconds or so. As soon as I increased the slider to 112% it went to 164W max then a second later it shut off as some hard limit was tripped, that's probably the only time I'll test that high in this P370EM.

Downside is the panel arrived and it doesn't work. The cable plugs in ok and it detects in device manager and the panel/res/modes/feature specs show up in HWInfo (LGD02C5 which is exactly as expected), but not one flicker from the backlight at any point. I'll try another vBIOS - maybe the external displayport-out not working is coming from the same source- the G65VR is such a weird custom beast it probably had its display outs (HDMI aside) routed through some custom board or some desktop-like PCIe link and maybe a proper laptop vbios with eDP stuff inside will fix that.

Getting pretty sick of pulling this laptop to pieces again and again. At least I can reuse the liquid metal - having to repaste every time would have sent me broke

 

Guesstimate correct.

Flashed 1BA1-1462-11FF v86.04.42.00.03 with a 151-170W tdp and... :

 

Hmm its code 43 after drivers install though. At least there are a range of vbioses I can try.

 

Log of the vBIOS roulette...

86.04.25.00.41 - 1BA1, 1462-11C2 [g65vr-1070] -both stock 115W vBIOS, and modded to 170W power limit
working HDMI out only (no DP out, no eDP out)

86.04.42.00.03 - 1BA1, 1462-11FF [gt62vr-1070] - modded to 170W
working eDP out. working HDMI out.
Driver install => device disable (code 43)
[couldn't test DP out without an installed driver]

86.04.56.00.3A - 1BE1, 1462-11FF [gt62vr-1070]
no eDP out
no HDMI out

86.04.2A.00.28 - 1BA1, 10DE-0000 [clevo 1070, p775dm?]
eDP out
no HDMI out
drivers = code 43

86.04.2A.00.27 - 1bE1, 10DE-0000 [clevo 1070, p775dm?]
no EDP, no HDMI
driver install [via teamviewer] fail code 28


NB.
I'm using an INF modded version of geforce 382.33
By "HDMI out", I mean to a passive adapter that plugs into external monitor via DVI-in. If a driver code-43's I can't change display outputs so can't explicitly test this. But the default is set to 'extended desktop' so it *should* display automatically if it's detected
By "DP out" I mean to same screen via DP-DP cable (does not happen without a driver installed so the code 43 drivers I don't bother testing)
By "eDP out", I mean internal display

 

Not good man. I'm sure you will have better luck with the P870DM-G.

Sent from my SM-T560NU using Tapatalk

 

I think the combo of inf mod/driver package was the issue. I downloaded the latest driver for the G65VR-6RE (375.70, the usual OEM cbf updating drivers for EOL models thing) and flashed an early vbios and it installed just fine!





In case youre wondering what "1070 SLI" is since this is only with one GPU installed, I changed the string at the end if the INF so I knew exactly what it was installing

 

Same driver, slave installed, sli active on internal display and dp-out, 86.04.42.00.03 with upped tdp works fine.

Not that I'll risk using 150W per gpu on this laptop for any length of time a power limit of 60% seems to roughly equate to stock tdp

 

i'm sure you can still OC to compensate.

 

I think an undervolt with a locked profile is how I'll go, locking to 0.85V/1700mhz was a good 10C cooler than letting it boost up to 1.06V. Downside is I'll have to switch manually to a profile that clocks back down (to 135mhz idle, 5W, ~35C) when not in game (vs idle 1700mhz/~50C)

Every bit of efficiency matters to me on a 5 year old laptop of unknown power circuitry specifications, I've had enough hard offs happen to question what is being tripped at times and the last thing I want to happen now that it looks like its all working is find a limit in a smokey and/or explosive fashion!

 

How long does it take to install a motherboard anyway?

 

lol ten minutes with all the practice I've had recently The real issue is getting a working one

 

I was referring to the P870DM-G mobo.

 

did you test it with unigine valley/heaven/superposition or 3dmark firestrike/timespy?

 

All tests run at CPU 3.9-4.1GHz, GPUs at +200mhz/+500mhz/85% power limit (at 100% util each GPU was around 125-135W and max temp was 79C master/85C on slave in Superposition). Max wall draw I saw was 365W (no CPU throttle since I put it on the dual adapter/joiner)

This is not max benching settings but is probably on the edge thermally. It was running dual screen with hwinfo/MSI AB graphs on the other, and with RTSS OSD. Running single screen and closing everything (readouts, steam, three cloud sync apps, AV etc) usually adds a couple %

Timespy 9111

Firestrike 18093

Firestrike Ultra 7929

Heaven 1080p/DX11/8xAA/Ultra/Extreme 3527 avg 140.0fps

Valley 1080p/Extreme HD 3561 avg 85.1fps

Superposition 1080p Extreme single GPU 3803, SLI 7498.

 

thanks

 

I envy you that you were able to upgrade your P370EM, lucky one!

If it would have been possible with the P570WM, I never would have sold it!

 

The old girl had a decent run at least.

 

@bennyg - Great work with the P370EM. I'm really interested to see how those 1070 SLI parts will work in the P870DM-G.

 

I don't think two of them will fit because of the tab section on the gpu.

 

Maybe not, but if it were done carefully, you might be able to shim the master GPU to raise it up just enough have the slave GPU tab go under the master GPU. Then very carefully tweak the heat pipes to accommodate the angle of the master GPU. I am reasonably sure it can be done if done carefully and the heat pipes are not kinked in the process. You would only need to shim it the thickness of the GPU PCB plus an extra millimeter or so to keep it from rubbing on the underside of the master GPU. Of course this depends on the heat plate fitting the GPU. Since @bennyg had no problem with the P370 heat sinks fitting, I am guessing the 980M SLI heat sinks will fit.

 

That was my plan. ~1mm or 2mm metal e.g. a washer under the screws will elevate the slave board above the master tab. The angle in the slot will only change by a couple of degrees. Bending the heatpipes to align the fit better is not new either I just have to be careful not to kink them.

Will have to cut the internal frame near the ram slots too.

 

Do the 980M heat sinks fit?

 

so you're thinking of lowering the mounting posts and bending the heatpipes to accommodate? that's just a crazy as sanding the PCB....I LOVE IT!

now to find 2 1070s with an SLI connector or get some ribbon cable and solder the wires directly.

 

I believe the 980M heatsinks cover almost everything needed to be cooled.

 

I am thinking that has to be the case since his original P370 heat sinks worked.

 

If I was *really* serious about not putting an XM in there I would have deleted the search notification off my ebay profile wouldn't I... and not seen someone selling a complete P370EM with 3940XM and a dead 7970M cheap enough to spend all of 5 seconds thinking about it.

Spare everything except slave GPU heatsink. I always get more adventurous when I have spares

 

dude take a deep breath and take another picture lol

 

Or get a tripod

 

Combo of a crappy phone camera that focuses on whatever it wants to, tapatalk not showing previews properly and me picking the wrong photo lol.

Its nothing special anyway, just a bone stock unit.

 

Basic results, CPU maxes at 4.3 (with the extra voltage and upped power limits), 4.4 gave an error on IBT. With liquid metal, maxed at 91C on IBT.

Firestrike went up quite a bit outside of the physics test: the higher frequency (and extra cache) increased the scores on GT1 and GT2 - total FS score 19,378 (up 9%). A GPU memory OC and maybe trying the ram at CL10 or 1800mhz should get 20,000.

http://www.3dmark.com/compare/fs/13343231/fs/13342093 (3940XM at 4.3ghz, vs the 3740QM at 3.9ghz, 4.1ghz single core turbo)

I also wasted a bit of time wondering why clocks were dropping in some of the tests even though I was increasing power limits... I forgot when I modded the vBIOS I dropped the temp limit to 85C. They're stable at 1950Mhz/0.95V/105% power limit which seems to run around 150-160W per card.

 

That's a nice voltage/core frequency match.

 

Finally found a way around an annoying cTDP/turbo power limit issue.

With the 3720QM/3740QM on both my P170EM and this P370EM previously, both with turbo power limits unlocked in the Premamod BIOS, I just set the short and long power limits way high in XTU and they set and stayed there, persisting everything, until a CMOS clear.

As soon as I put the 3940XM in, on boot, it kept on setting cTDP profile 2 for a 81.25W short and 65W long turbo limit. There are zero BIOS overclocking options enabled in the BIOS (Premamod 1.02.09/1.02.08). XTU would set the multis correct but the power limits would not be applied on startup - to change the turbo power limits I had to do it manually through the XTU GUI, which would try and tell me the power limits in the profile were currently loaded, lucky I had HWInfo to tell me otherwise and clocks would drop to as low as 3.7ghz on an extended bench, hitting the 65W long power limit. Unacceptable.

Changing Throttlestop settings on the TPL page wouldn't change anything at all. I have it to run at startup so it sets the turbo multipliers (x43) as normal.

So I found the command line way to use XTU - XTUCli.exe - and using a short Powershell script it runs automatically on startup and sets the Turbo Long power limit to 95W and the short to 90W:

Here's the script: (adapted from https://www.reddit.com/r/Surface/comments/3vslko/change_cpu_voltage_offset_with_intel_xtu_on/)

Code:

$status = get-service -name "XTU3SERVICE" | Select-Object {$_.status} | format-wide
if ($status -ne "Running") { start-service -name "XTU3SERVICE"}
& 'C:\Program Files (x86)\Intel\Intel(R) Extreme Tuning Utility\Client\XTUCli.exe' -t -id 48 -v 90.0
& 'C:\Program Files (x86)\Intel\Intel(R) Extreme Tuning Utility\Client\XTUCli.exe' -t -id 47 -v 95.0
sleep 5
stop-process -id $PID -force

Set it to run via Task Scheduler (very similar to the standard Throttlestop task scheduler setup)
General tab:
Run only when user is logged on
Run with Highest privileges
Triggers tab:
1) At log on of any user; Delay 20 seconds
2) On event: Microsoft-Windows-Kernel-Power/Thermal, Microsoft-Windows-Kernel-Power, Event ID 507, Delay 20 seconds [event 507 is a resume from hibernate you could add one for a sleep resume - I don't use sleep, and anyway in a couple of tests it seems sleep crashes the video drivers on waking]
Actions:
Program/script: powershell.exe
Add arguments: -ExecutionPolicy Bypass c:\pathtoyourscriptfile\filename.ps1 [the bypass executionpolicy switch may help solve administrator issues]
Conditions: all unchecked
Settings: all unchecked; tick 'allow run on demand' for testing the script

The startup delay is to give the system time to start the XTU services, and also give a window to stop the script if somehow it starts borking up (unlikely with power limits but that's a possibility if you set an undervolt or a typo which BSODs as soon as it's done)

You could automatically set your whole XTU profile to load using these commands if you wanted, but of course TS is the preferred program for everything that it is possible to do.

Open cmd.exe as admin and navigate to your XTU install directory (default C:\Program Files (x86)\Intel\Intel(R) Extreme Tuning Utility) and under "Client" subdirectory you'll find XTUCli.exe
Run "XTUCli -i ALL" to output a list of all the parameters you can alter on your system. The ID of the parameter is what you're after - in my case 47 for turbo power short and 48 for turbo power long. I could set them to crazy high numbers but 90W is enough for now as the highest I've seen on this CPU was a tick over 80W and that was Intelburntest at x44 just before it errored out.

 

OS in Legacy mode: XTU writes values into BIOS.
OS in UEFI mode: XTU needs to load them from within OS.

 

Yeah I don't think UEFI likes you trying to write straight to the bios

 

I honestly don't know what the uefi is set to but the less i have to use of bloated xtu the better.

CPU was able to encode stable at x45 at 1.41V (+100mV extra), for less than a minute due to temps. x43 at 1.296V tops out low 80Cs (low 70s in games).

It also absolutely monstered through some video editing in Davinci Resolve. On selecting a time point I'd have to wait a couple of seconds or so for previews to generate on the 3740QM @ 3.9, on this CPU it just did it all in real time, no stuttering at all.

edit: lol frequency/volt wall.
TS Bench 32M
at x43 (1.296V, +0mV turbo): 64-65W, ~7.1 sec
at x44 (1.33V, +50mV turbo): 69W, ~6.9 sec
at x45 (1.40V, +100mV turbo): 79W, ~6.7 sec

 

It is beautiful. Nothing else has to be added.

 

FWIW it's set to legacy

Maybe that's the issue. I switched my P170EM to UEFI when I installed Premamod BIOS and switched to Win10 (which broke needing reinstall) and afterwards the change in turbo power limit has stuck and not needed to be changed in BIOS ever since

A couple of things...
- driver update
- raise temp limit on vbios - the fans don't kick in to highest until something - usually slave GPU - hits 90C, then it'll settle low 80s - however since I set the hard temp limit at 85C it just reaches that and throttles down to 1600MHz-ish bouncing off the 85C limit, unless I do max-noise fan mode
- some A/B testing on my bottom case mods, very easy since I have an unblemished bottom case as well. Seems to be working well for master GPU (it sits around 70C, limited to the same clock as the slave) - but not so well for the slave.

 

Not having any success with drivers newer than 375.70.


Also, 858 XTU marks is 10 points off the highest P370EM XTU score - and mine is with budget 1600CL11 DDR3 ram. I have some 4x4 2133 that I found cheap that arrived yesterday ready to go in so that should make this officially the fastest P370EM ever

http://hwbot.org/submission/3641262_

x45, 1.41V (+98mV extra turbo voltage), max fans
Other results:

Thermal throttled a bit on TS Bench 1024M.
32M 6.58sec
1024M 207.8 sec
7zip 3517MIPS/thread, 25553MIPS total 738% scaling

CPUz bench : 451.8 single / 2241.6 multi (vs a stock 6700K: 474ST / 2377MT)

It will thermal throttle hard under AIDA64 FPU so I'm not going to bother.
Intel Burn Test: 89GFlops. but... Multis were dropping to below x42 and I realised it changing the PL hadn't worked despite HWInfo showing it had, and it was still stuck on a 90W turbo power limit ... maybe hard coded?

Total liar ;p

 

I delivered P370EM3 1070 in the notebook, but I want that it worked on the internal display but not through the monitor from the PC (It is possible?

 

I have all displays working, internal display, and hdmi and dp out.

Have not yet got a driver later than 375.70 working but I realised recently I was downloading the *notebook* nvidia driver packages instead of the *desktop* packages that the inf mods are for. So something else to try.

 

You'll have less issues with the P870DM-G + MSI 1070 combo.

 

Wow. I think I know what my next mod is going to be. Do you forsee any potential issues with getting this to work with an internal EDP 4k panel?

 

I lack confidence in my ability to accurately repin the w230ss qhd cable sitting in the draw... that's my issue. I think the 30pin edp is wired the same across all the p370 refreshes

 

Hi everyone!
Bennyg,you've done really great job!) If you can,contact me ...I got a little problem with my 1070m and one of your photos can save me ^_^
Need name of this component,which is bobled on my photo....

 

What happened to it?

 

Suddenly laptop reported about crash of video driver and after reboot there was only one GPU in device manager. After disassambling i've found this thing.

 

Taobao card, or some other source?

I kinda want to jump to a 1070, taobao is a bit cheaper, but warranty or even assurance that is not dead is not great, and from ebay, its 1000€

 

I'll grab a super close up pic for you this weekend. I have to reflash the vBIOS back to the stock vortex one, literally no other driver or mod I've tried other than the MSI 375.70 package works. UEFI made no difference, god dammit reinstalling Windows (after converting storage to GPT) is such a hassle.

Got my hands on AOTS and there's something seriously bottlenecking SLI(DX11)/mGPU(DX12), the GPUs sit on about 40-50% utilisation even downclocking themselves, and get around the same fps as a single 1070. The AOTS benchmark util says its >99% GPU limited (compared to 100.0% in single 1070).

One thing I noticed in GPUz is in the NVIDIA SLI field at the bottom it says "high-bandwidth", is this normal for an old SLI cable dating back to the Fermi era?

 

I've wondered about this since the Pascal laptop cards came out. They only have 1 SLI connector per card (and it is the same connector used on legacy gpu's), so they are not using a double bandwidth SLI bridge.

GPU-Z reporting "High bandwidth" could simply mean that it's detected a SLI Pascal configuration and it interprets that as meaning it must be High Bandwidth; i.e. it's not actually reading the real SLI bandwidth numbers.

In addition to using two connectors, SLI High Bandwidth uses a pixel clock of 650Mhz versus 400Mhz for normal SLI. Perhaps Pascal laptop cards running in SLI are using 650 MHZ pixel clocks and that's why it's called high bandwidth (however the bandwidth will still be half of what is available to desktop cards using two fingers.)

Are the Pascal laptop SLI bridges constructed any differently than the Maxwell laptop SLI bridges? Perhaps your bridge is not capable of running at 650Mhz and thus limiting performance?

However, from what I can tell, the SLI bridge bandwidth is only a function of the number of gpu's, the resolution, color depth, and framerate you are pushing. If you are running a 1080p screen at 60hz or 120hz, you should not be limited by a 400Mhz single SLI bridge. The way I've interpretted the information (and I'm probably wrong) is for 1080p at 60hz with 32bit color depth, the SLI bridge would have to be running at roughly 124Mhz.

For 2160p at 60hz and 32bit color depth, the SLI bridge would have to be running at roughly 495Mhz

I'm currently running two Maxwell mobile chips in SLI and they have no utilization issues pushing 3820*2160*32 at 60hz. That means, from what I can tell the SLI bridge used for Maxwell mobile cards is running at at least 500Mhz.

Nvidia has something called LED SLI which runs at 540Mhz. Maybe this is what is used on Maxwell mobile cards.

So assuming you're running at 1080p and 120hz you need to have a SLI bridge capable of running at ~250Mhz. The fermi era bridge should be capable of running at at least 400Mhz so I do not think this is your bottleneck (unless you are getting the utilization issues when pushing much higher framerates and resolutions).

Or it could be that the cards (or software) are detecting that your cable is not capable of High Bandwidth SLI (i.e. 650 Mhz, 32 bit) and pre-emptively limiting the SLI bandwidth to ~half of what it should be?

I'd be curious to hear what someone with more knowledge has to say on this. From what I can tell, the only reason you would need to have SLI High Bandwidth (specifically with a 32 bit bridge instead of a normal 16 bit bridge) would be for pushing insanely high framerates and resolutions like 4k at 120hz. Lower framerates and resolutions should not be affected and perhaps this is why the mobile Pascal cards only use 16 bit bridges? (But are they running at 650Mhz?)