Clevo P750ZM: GTX 980M Overclocking Investigated
by Jarred Walton on March 20, 2015 10:00 AM ESTGTX 980M Overclocking: Do the Time Warp
With the overclocking results in hand, it’s not too difficult to see why people would be upset with NVIDIA locking out overclocking on mobile GPUs. Using the Clevo P750ZM from Eurocom (aka the Eurocom P5 Pro), we were able to easily increase clock speeds on the GPU by more than 10%, and a 20% boost only required the use of an unlocked VBIOS. If NVIDIA follows their usual pattern, the overclocked GTX 980M is likely going to be faster than the next halo mobile GPU from NVIDIA, whenever that shows up (probably in the fall).
Hefty, hefty, hefty...
With those results in hand, NVIDIA’s decision to lock down mobile overclocking might seem like a way to prevent users from getting a free performance boost, but there’s more going on. The biggest concern is that we’re using what is arguably one of the best platforms for GTX 980M overclocking in this article, and while the results are great I wouldn’t count on all notebooks using GTX 980M being able to achieve similar clock speeds.
Case in point: the Gigabyte P35W v3 discussed on the previous page. Depending on the game and settings, the CPU can be a bigger bottleneck than the GPU, and particularly at lower resolutions (e.g. 1080p) the GTX 980M needs plenty of CPU power. If a CPU has to throttle down to lower clock speeds due to excessive heat, overclocking the GPU isn't going to help much. Our stress testing of the P35W v3 seems to have encountered this scenario, and there are times when just running a game on its own can trigger CPU throttling.
Thin + overclocking = questionable
As with any overclocking attempt, the final stable limit will vary even with the same core hardware. We managed a 20% GPU overclock without too much difficulty on the Eurocom P5 Pro (P750ZM), but other P750ZM notebooks might do better or worse than our sample. The Gigabyte P35W v3 with the stock VBIOS also managed to max out the GPU overclock (around 12%), though the benefits aren’t as great thanks to the slower CPU and less robust cooling.
The big question is going to be what overclocking might do to long-term stability and component life. Running the fans in the P750ZM at 100% keeps the GPU running nice and cool even when overclocks, but it could cause the fans to wear out sooner than expected. Having the CPU run in the mid-90s is also a concern, so you might consider buying an extended warranty (even if you don’t plan on overclocking). And while we didn’t have any problems with our P750ZM during testing, we’d be remiss if we didn’t note that we’ve had more than a few high-end gaming notebooks go belly up over the years.
Only NVIDIA and their partners know the exact numbers, but it’s possible NVIDIA has been seeing unacceptably high numbers of mobile GPU failures and locking down overclocking by default would certainly reduce the number of people willing to pursue that route. The true enthusiasts meanwhile can still get what they want by purchasing notebooks that aren’t locked down (e.g. the Clevo P750ZM). Just be careful how far you push the clocks, though, as if you burn out a mobile GPU like the GTX 980M, buying a replacement can cost as much as a desktop GTX Titan X GPU.
There’s a corollary to the overclocking discussion that’s also worth mentioning: underclocking. Sadly, we’ve seen more than one gaming notebook over the years where running stock clocks proved to be difficult if not impossible with certain games, particularly once the notebook is a year old or more. Mostly that has been with mainstream GPUs (e.g. GT 750M and similar), and underclocking by 10-20% will typically fix the problem. We’ve seen others mention this problem as well, and if NVIDIA fully locks down the clock speeds on mobile GPUs it could create unintended consequences for those trying to underclock.
The final tally is that after backlash from the buyers, NVIDIA has decided to not lock out mobile GPU overclocking in their drivers – and true to their word, the latest 347.88 beta drivers allow mobile overclocking again. (The P35W v3 OC testing was done with those drivers.) We might see NVIDIA try to limit overclocking in the future, but hopefully they stick to using a locked VBIOS. That would allow the notebook manufacturers to choose whether or not they want to support GPU overclocking, and that’s probably the best approach as ultimately they’re the ones dealing with RMAs.
And let me end with this warning: I have burned out GPUs in notebooks in the past, even without overclocking, doing stress testing. It’s a terrible feeling to have a notebook suddenly refuse to boot, or crash as soon as the graphics drivers load. While getting a “free” 20% boost to frame rates is definitely nice, replacing/repairing a dead notebook is not. Discretion in this case is not a bad thing, and overclocking a notebook GPU just because you can isn’t without risks.
Facebook
Twitter
RSS
34 Comments
View All Comments
JarredWalton - Friday, March 20, 2015 - link
Well, -100mV is *trying* to run my stress test. It crashed without max cooling (just Tomb Raider, not the system). I maxed the fan and tried again; after a minute or two the system restarted. Guess I'll try -50mV now.WJames65 - Friday, March 20, 2015 - link
I bought a barebones Clevo P770ZM equipped with a 980M and added a 4790K. The 4790K seems pretty receptive to undervolting based upon my experiences and experiences of many others as posted in the TechInferno forums. Some people have managed to get as high as 5 GHz with 4.7 GHz more typical. However, for such high speeds, the fans must be run on max, which is accomplished with the hotkey FN + 1. I also suspect those higher speeds may require a delid *shrug*. While I personally wouldn't run a desktop processor in a laptop with those overclocks, it is possible.JarredWalton - Friday, March 20, 2015 - link
Are you saying people are hitting 4.7GHz under a stress test? Because I can get the system to boot at 4.6/4.7GHz but as soon as I start stress testing it's a no go. Trying for -50mV on the CPU now as -100mV didn't work out.ajc9988 - Friday, March 20, 2015 - link
So, take it as you will:I recommend starting with -25mV at a time. First, lock the cache ratio multiplier to 40 (hereinafter stock multiplier). Start at -30 to -50mV under on the cache ratio (note - leave all adaptive voltages set to adaptive and default. If you try setting the adaptive voltage like a desktop, then set an offset on top of it, some weird readings can be had in programs like HWmonitor for both the ring voltage and the IA voltage offset. As I haven't figured this out yet, I'd recommend staying away from it until later if you want to play). Then reduce the voltage from that amount by 25mV until no longer stable in XTU stress test. Then bring it up by 5mV intervals until stable for the test of your choice. What you may want to explore is other tests utilizing the AVX2 instruction set as this is the cause of the extreme heat on Haswell with the newer prime95. Alternatives for lower heat tests are prime95 version 26.6. If you want AVX ( but not AVX2) you can use prime95 27.7 or 27.9. These give you prime95 stability without the insane heat for the platform. Alternatively, loop x264 for testing the AVX set. Either way, the idea is to have the laptop under normal stress situations, not insane heat situations unless you do certain encoding that requires it. Then it is necessary. You just want fully stable. After you find the stable cache ratio, repeat for the core clocks. Lock the multiplier for all cores to be the same and start offsetting. (note - if the C states are enabled, dynamic voltage offset must be true for every possible C-state voltage request. Think of it like a totem pole with set notches in it. These notches correlate to stock voltages for every C-state, including C0 (load). The offset is like cutting that much voltage off the bottom of the totem pole. This means instability may arise entering or coming out of C-states. Just be aware of it). Once you have the Core and Cache offset stable to your liking, then test away with what you like. Benchmark, do your every day tasks, etc. Be aware, it may pass any number of stress tests and still have slight instability. Just increase the Core voltage and keep going. As long as you are under 4.4 ghz, you shouldn't have a need to change the cache voltage for stability (in my experience with my machine, yours may differ). I run everyday at undervolt -65mV Core, -110 Cache (0.995 Vcore) at 40 multiplier all cores and stock cache multiplier. That is low 70s auto fan for XTU and WPrime testing. After I get my pads here, gelid gc-03 extreme, and CLU, I intend on delidding and going beyond (the 4.9GHz was on a delidded CPU). 4.2 should be easy on max fans also in the 70s on lighter stress testing (meaning other than prime95 v.28 or OCCT 4.4.1).
nunomoreira10 - Saturday, March 21, 2015 - link
Did you manage to undervolt? any positive results?ajc9988 - Saturday, March 21, 2015 - link
Also, first try to get 4.5 stable in lower stress test conditions first, then see if the thermal room will allow for 4.6 or 4.7. Unless you won the Haswell chip lottery, 4.6 and 4.7 without delidding will not likely be possible, but up to 4.5 should be if you don't mind seeing temps in the 90s. I am delidding solely to get 4.4-4.5 under autofan in the 70s. I, personally, apply two different settings: work mode and play mode. Work mode is autofan and temps in the 70s. This leaves plenty of thermal headroom and allows for it to be quieter in public, etc. The other is game mode. This is 70s to low 80s at more stress than I do for work mode. It is important to remember that this is a desktop chip with a Tcase of 74.4 degrees Celsius. Laptops are extremely closed cases, meaning that the heat around is nearly what the CPU temp is, meaning 80 should be your upper max for everyday operation (or at least is how I treat my chip for longevity). You will find different offsets on Dynamic Voltage Offset (Vcore offset) depending on the multiplier used. This is because Intel guarantees all chips will function (up to max turbo 4.4GHz) at the stock voltage applied for each multiplier. Each chip is unique and some can use less voltage at a specified multiplier. So the offset will vary for the Vcore depending on the multiplier.Also, there are three types of overclocking: Static, Adaptive, and Dynamic Mode Overclocking. Dynamic Mode Overclocking is what is commonly applied to laptops. This allows for lower voltages while allowing for programs to request more than what the offset normally applies for voltage without crashing the system. Benefit - lower overall voltage (meaning lower heat). Drawback - stability testing sucks, C-states are affected by the voltage offset. Static - applies a set voltage all the time. Benefit - stable under any draw. Drawback - Voltage applied all the time which increases overall heat, requires much higher voltages than any other method of overclocking. Adaptive Overclock - Adjusts only the turbo voltage requests on the CPU. Benefits - Does not affect the voltage of C-states, it only affects the Turbo Voltage (acts as a separate VID state above all C states); allows CPU to request voltages beyond the max turbo voltage set by user. Drawbacks - allows CPU to request voltages beyond the max turbo voltage set by user which causes extreme power draws when running synthetic benchmarks like Prime95, OCCT, etc. Because of these, Dynamic Mode Overclocking is the preferred method of overclocking on laptops. But see: http://forum.notebookreview.com/threads/guide-to-h... . I have not been able to replicate his findings/results with adaptive overclocking while performing Dynamic Mode Offset. Further, I run my CPU under stress at a dynamic offset causing the Vcore to be .0.995V, whereas with setting the Adaptive voltage to .950 Vcore, it still supplies 1.055V Vcore under stress conditions (as with the heat that comes with it). This may be different once the voltage required is greater than Intel's stock voltage at a set multiplier, but for voltages used before that point, it would seem that Dynamic Mode Overclocking provides the lowest voltage to the CPU, and thereby provides the lowest possible heat for the 4790K without advanced cooling solutions (delidding, etc.) for this device.
extide - Saturday, March 21, 2015 - link
Yeah those guys seeing 4.6+Ghz on the laptops, are definitely no stable under prime95 and such.ajc9988 - Sunday, March 22, 2015 - link
Not true. 4.6 and 4.7 was prime95 v.28.5 stable. See http://forum.notebookreview.com/threads/official-c...Now, this immediate testing was XTU, but prime testing did happen (see page 255). Superkyle1721 continued his delidded OC journey to page 258. I, personally, have prime tested and had my web surfing (I'm a tab whore) BSOD my system on Prime stable. This means prime stable isn't the be all end all when firefox brings prime stable to it's knees (over 300 tabs at a time is normal). So go on with prime, it is hot and doesn't guarantee stability for my use. I look at prime95 27.7/27.9 for my AVX testing (10 degrees hotter than others rather than 20 tested by AVX2 instruction set found in prime95 v.28.5 and OCCT4.4.1.) as well as x264 loop. I start with XTU because it is about the easiest to pass. Then move onto Wprime and BurnIn Test (Passmark). After it is stable in these, I do an x264 loop. Then I do either prime95 26.6 (no AVX) or prime95 27.7/27.9 (if testing AVX and not AVX2). Then I do massive blend testing. Heat is not what tells me I'm stable because a pass in prime still fails my browsing habits, needless to say more of my visualizations. It is a great tool at pointing out instability - I admit it is one of the best. But choosing something that is one of the best at what it does that has been made impracticable by heat considerations is ludicrous! If you want a full list of my stress testing programs and benchmark programs, I can get that to you. But Prime95, due to its heat with AVX2, has made itself impractical for most uses and few people do activities that push their processor to those limits. So the average overclocker is building in a 20 degree headroom on his/her system that he/she will never take advantage of. Does that make sense? NO, IT DOES NOT!!! Some people encode, run multiple virtual machines simultaneously, etc. and push their processors to the max on temp reaching the same as Prime95 small FFT. Those people have a need for this to set their thermal limits of their systems to their liking. But for the rest, you set prime95 small FFT to 80 and never see the processor break 65 under load. Why? It is time to start rethinking the Prime or nothing standard. Yes, testing with multiple stress tests to assure stability takes longer. But you can set the temps according to your usage, not some theoretical max that you never push in your use of your machine. Believe it or not, I have had 4.2GHz prime95 stable small FFT with thermal throttle to 4.15 with bad heatsink contact on one of four corners with MX-4. I have had it 4.5 stablish with thermal throttle to 4.42 with XTU and WPrime (I will not prime test if weaker tests thermal throttle, that is ridiculous to consider and you should have your head examined if you think otherwise). Superkyle1721 had, after delidded, got 4.4 prime95 for 20min. large FFT with temps never above 92 degrees. Granted not as hot as small FFT, but still a fair score. So please read and research before saying those guys are not prime stable as if that is the only stability that matters. Open your mind to new possibilities please.
scook9 - Friday, March 20, 2015 - link
The problem with desktop CPUs in a laptop is that the cooling is WAY less efficient due to the use of a heatspreader. Laptop CPUs have the exposed die with no heatspreader in the way so you have a much shorter and more efficient path for heat to flow and cooling to be effective. Combine that with the fact that the laptop CPUs are binned for much higher efficiency than desktop parts and it starts to explain why laptop CPUs make more sense in laptops :) They do cost more as a result, but pretty much since Sandy Bridge, there has been very little reason to bother with desktop parts in laptops, clock for clock the performance is identical.An Intel Core i7 4810MQ can turbo up to 3.8 GHz vs 4 GHz for a 4790 or 4.4 Ghz for the 4790k. At the same time power usage is HALF resulting in much better thermals and more reliable performance as well as gaining the benefit of technologies like Optimus. If you are really that hell bent on CPU performance for gaming in a laptop than money is clearly not a limiting factor so get the 4940MX can clock it well past the stock clocks of a 4790k (this is very doable). At this class of laptop the premium for the extreme CPU is not as bad as it sounds, adding $800 to a $3000 laptop is doable for the people buying $3000 laptops usually. I have played that game in the past, getting the extreme CPU and overclocking the hell out of it in a top tier gaming laptop and after getting benchmarking out of my system (which took about 3 years) I now usually go for the entry level quadcore to have all the threads I want and not pay out the nose for marginally higher clocks.
What I was able to pull off back in 2011 (with ambient room temp cooling - some guys were doing phase change and dry ice):
http://forum.notebookreview.com/threads/official-a...
Got that CPU up to 4.5 GHz pretty reliably on all 4 cores with no throttling.
scook9 - Friday, March 20, 2015 - link
I also was able to push the hell out of the older Core 2 Quad era stuff too:http://forum.notebookreview.com/attachments/3dmark...
http://forum.notebookreview.com/attachments/3dmark...
But that DID require putting my laptop in a chest freezer to cool the CPU - 1.5V ran pretty hot. At the end of the day though, the crappy nvidia chipset in the original M17x was my limiting factor with my Ram and not the CPU.