since I spent around 20 hours in the last four days with benchmarks, testing and thermal issues, I have now a great setup which I would like to share with you.
Problems I ran into while the machine was under heavy load (e.g. Fortnite, rendering, Cinebench, AIDA64, etc.):
Problem1: After a few minutes, the CPU clock was reduced to the minimum clock speed of 0.78 Ghz. Apart from that, the system didn't use the Nvidia GPU anymore but switched to the Intel GPU. The system was now barely usable. After around 5 Minutes, everything went back to normal.
Solution1: After a long forensic work and forum reading, I managed to identify the "Intel Dynamic Platform and Thermal Framework" as the cuplrit. Uninstalling it did not go well (it keeps coming back, no matter what you try...). It seems like some Intel Software updated the Thermal framework to the newest version, which renders your machine unusable. To solve this problem, download the Intel Thermal Framework from here: (edit, new link to DPTF drivers Version 8.3.10207.5567, A09: https://www.dell.com/support/home/de/de/dedhs1/drivers/driversdetails?driverid=d4g6r - just confirm the prompts "the driver will be replaced by an older version" stuff. You can prevent the newer (faulty) drivers from being installed e.g. via a group policy (windows 10 pro): https://www.howtogeek.com/263851/how-to-prevent-windows-from-automatically-updating-specific-drivers...
There are different types of throttling; a) thermal throttling (CPU/GPU/VRMs) are getting too hot and b) power limit throttling (the machine has a power limit which can't be exceeded by a single part or the machine as a whole). The i9 of the XPS has a power limit of 56W permanently (please correct me if I'm wrong), or 78W (short turbo boost period). You cannot do a lot against the power limit throttling (which is good), but thermal throttling can be solved completely with this machine.
Problem2: CPU temps were not too bad out of the box. The problem was that bumping up the speed of the CPU at the beginning of heavy load, lifted up the temperatures of the cores to >95°C and the machine began throttling in an instant. Temperatures recovered after the fans kicked in and it stayed around 85°C to ~91°C with nervous thermal throttling sometimes.
Solution2: I repasted the CPU/GPU with another thermal compound (IC Diamond 24carat). Temps are now going up much slower, even without thermal throttling at all. The cpu frequency of the machine is now solely limited by the power limits of the VRMs. Screenshot:
You can see that the package power is constantly hitting the 56W mark (right top) and the CPU frequency is extremely stable. The small temp spike in the beginning (~78°C) is due to the fans which need around 10 seconds to kick in. GPU stays at ~75°C under heavy load in the package (CPU and GPU at full load). Fans kicking in a lot later and not with such a heavy spin. Fans are completely quiet when browsing/office work. I can even hear the crappy
WD15 TB16 dock
In addition to that I experimented with undervolting (ThrottleStop and Intel XPU (*edit: use ThrottleStop only)). Undervolting the CPU/CPU Cache at -80mV-100mV brings the CPU temp down even more and you can get a bit more clock frequency at the same power limit (e.g. AIDA CPU only test before: 3.52 Ghz (no temp throttling, power limit 56W), after: 3.82Ghz (no temp throttling, power limit 56W))
Here is a video how to load ThrottleStop on startup in "silent" mode: https://www.youtube.com/watch?v=K1srWgovdn8
Benchmarks after changes:
Cinebench R15 (only first runs, no loops): 1270-1320 (multicore) / before it was around 1020-1080.
CPUMark (after changes): 15600-15750 / before: not tested
ThrottleStop: Disable CPU turbo (CPU power consumption and its induced high temps only obstruct the GPU from performing well)
In game: cap framerate to 60 fps, I play it with all settings to high and view distance to epic on constant 60 fps (which is enough for me). The machine power intake will never go far beyond 100W which is not a problem for the machine's thermals.
Dell XPS 15 (9570),
Intel Core i9-8950HK
Windows 10 (stock)
BIOS version is 1.2.2.
After figuring out all this, I'm now very happy with this machine!
Disclaimer: You have to know what you're doing! Repasting is not an easy task on a brand new machine! For ThrottleStop, there are many tutorials out there.
Solved! Go to Solution.
Did you check the temperatures in the "Dell EC" group using HWinfo64, in particular the first Ambient and DIMM? (Mind the sensor labels may not be the same in Aida)
> You cannot do a lot against the power limit throttling (which is good)
I don't think throttling is good. It may not be avoidable with the i9 9570, but it can likely be reduced.
VRM heat is the main problem. Did you pad the VRM mosfets? iunlock's mod for the 9560 seems to be the best known solution, but it is a bit involved.
After you take care of VRM heat, you can consider padding the RAM, particularly the DIMM sensor in between the sticks that appears to trigger PL throttling at 63 in the 9570. To the backplate (but assuming you haven't padded VRM mosfets to the backplate). This should reduce PL throttling.
>>Did you check the temperatures in the "Dell EC" group using HWinfo64, in particular the first Ambient and DIMM?
Yes, AIDA is not as good as i thought in the first place. So now everything is being monitored with HWiNFO64.
I tested multiple mods out there - padding the VRMs / Mosfets to the backplate had the most success. The heating of the backplate is not heating up the fan's air too much because the surface that comes into contact with the air is fairly minimal (and additionally cools the backplate).
>>I don't think throttling is good.
Right, let me put it this way - a slim laptop consuming more than 130W over a longer time will glow in the dark. So with the current 14nm++ technology, this kind of laptop should not use more power (at least with this kind of cooling technology).
I will try padding the DIMMs - the backplate in this area is fairly cool even though the VRMs are heatbridged to it as well.
> I tested multiple mods out there - padding the VRMs / Mosfets to the backplate had the most success. The heating of the backplate is not heating up the fan's air too much because the surface that comes into contact with the air is fairly minimal (and additionally cools the backplate).
This is known to help a little bit, though not so much. But when folks tried better Fujipoly pads, the long-term performance deterioration has shown up. It is okay if using a laptop cooler stand, but actually worse without.
> Right, let me put it this way - a slim laptop consuming more than 130W over a longer time will glow in the dark. So with the current 14nm++ technology, this kind of laptop should not use more power (at least with this kind of cooling technology).
iunlock's mod was successful in making the i7 9560 run unthrottled. It took him quite a lot of effort to come up with, and still takes an effort to repeat. But it shows it is feasible in the same space (not sure if also for the new 6-core CPUs, despite the same TDP in the i7). For manufacturing, cooling would have to be redesigned of course. Looking at some other laptop designs, they could extend the radiators a little bit, and add another small heatpipe for the VRMs.
so a lot of people have recommended to pad the VRMs (Mosfets too?) to the backplate of the 9570 (not only the 9560 / 9550, right?). But what I haven't seen so far is what chips exactly to put the pads on? Any ideas of even photos of this? Would be much appreciated, thanks!
I keep getting the problem on my XPS 9570 i7 8750h, I did all of your steps but it fixes it for a bit and the problem keeps coming back, When I run Cinnabench R15 my score is in the 400's but when I first received my xps two days ago i was getting 1,200's
In the 400s? That's very low. Sounds like your CPU is throttled to a minimum. Can you open the Intel Power Gadget while your Cinebench is running to keep track of your clock frequency?
could you report what CPU clocks you get initially and what after 15 min of the Prime95 stress test (blend)
a) with 12 threads
b) with a single thread?
and after how many seconds PL throttling kicks in?
It appears the i9 power limits are locked down in BIOS and that its long-term turbo clocks (which depend on the number of cores loaded) are lower than some other machines running an i7.
- initially, the clock frequency is around 3.7-3.9Ghz for around 32 seconds (then PL1 kicks in)
- after around 20 Minutes, the clock frequency is stable at around 3.1-3.2 Ghz (PL1 56W)
- after around 20 Minutes, temperature is constant at around 79-81°C
- after around 20 Minutes, total system power consumption is constant at around 95-97W
- initially, clock frequ. between 4.1-4.6Ghz (Package Power around 47W)
- stays like that after 10-12 minutes, frequency fluctuating (3.8-4.6Ghz)depending on specifics of the performed test
- package temperature around 80-86° (sometimes dipping higher)
- no thermal throttling but PL4 (Max VR Voltage) is Yes in HWiNFO
Thank you. Intel doesn't seem to publish a table of max boost frequencies vs the number of cores anymore, but long-term 3.1-3.2 Ghz is less than what some others get with i7-8750H based laptops (i5-8300H report even higher clocks but these have 4 cores not 6).
It's hard to compare those. CPU frequency also depends on workload (tested chipset command branches). i5 may have to ramp up more Ghz with a higher load overall than an i7/i9 would.
In all core cinebench for example, clock speeds are higher while all cores being under 100% load.
> In all core cinebench for example, clock speeds are higher while all cores being under 100% load.
Yes the clocks in Cinebench are higher, I guess the CPU load is intermittent not so continuous as in Prime95
>In addition to that I experimented with undervolting (ThrottleStop and Intel XPU (*edit: use ThrottleStop only)). Undervolting the CPU/CPU Cache at -80mV-100mV brings the CPU temp down even more and you can get a bit more clock frequency at the same power limit (e.g. AIDA CPU only test before: 3.52 Ghz (no temp throttling, power limit 56W), after: 3.82Ghz (no temp throttling, power limit 56W))
Wait hold on. Are you sure you are running at the 56W PL1 limit when doing 3.82ghz? I've basically done everything you've done and yet my 56W limit is still 3.5ghz. Can you confirm with throttle st
Ah, I see. Yeah I've been running cinebench with stresses the whole CPU. That makes sense because not stressing cache and fpu reduces power consumption and increases my clock speed. I really wish dell allowed advanced bios editing so I could increase the power limits. I have so much headroom and no ability to utilize it.
installing this framework, immediately fixed my 9575, and it began to behave like it should. before that, it was stuck mostly at 0.8GHZ power limit...THANKS!!!
I have a Dell Latitude 5591 with i7 8850H and MX130. I have an external thunderbolt docking station hooked up with 2 4K monitors. I have been fighting this crazy crash throttling for a week. Machine runs great under normal use. Clock goes up and down slightly. When I try to watch a stream video I get the clock crash to 800mHz. I have been trying different setting with throttlestop and XTU and windows power management while monitoring speeds, voltage, and temps with several programs. Nothing showed terrible issues. Throttlestop would show a PL1 throttlestop when the clock would crash out but that’s about it. The PL1 has me thinking power limit throttling and XTU showed that same but why throttle to 800mHz on a 4000mHz processors? Undervolting would leave the cpu temps sub 80c but still the crashes. I lowered the resolution from 4K to 2k and that helped but that is a sticky fix. I tried using then imbedded intel graphics and then the MX130. Nothing helped. I finally found this article and some from Apple users complaining about VRM throttling. I could never find a sensor reporting my VRM temp but it sounded like the same issue. I tried several intel thermal framework versions but all behave the same on my 5591. Finally I got a $20 laptop cool pad with 3 fans that blow up through the case. Solved all the crashing. I can now watch a YouTube or twitch without the crash out using stock settings. It looks like it is really all about a cooling issue with the VRM.