A comparison between Clevo mobile processor overclocking and a 4770k

Hey Folks, I want to have a conversation about why different people are having various levels of success when it comes to overclocking mobile processors. The motivation for this thread is to gather some different experiences from other users with regards to what methods are working with which machines. I also pose a question as to why desktop processors perform so much better, at least in my case. If this interests you, bear with this epic series of posts. If you prefer TL;DR, it’s in the second post.

First, I love having a TON of CPU power in my primary computer. My office setup is as follows:

- Desktop or laptop-of-the-week (I joke, but seriously...this week it’s a P650se. Next week it'll be a P750zm)
- 3x 1080p LG IPS monitors (Connected via HDMI / Displayport interfaces)
- 1x 2560x1080 LG IPS monitor (Connected via a USB 3.0 to Displayport Adapter)
- 1x Displaylink USB3.0 Docking Station (Daisy-chaining DIsplayport Adapter)
- Various USB Devices (Fiio DAC, External RAID Array, other External HDDs, etc)

My typical relaxation scenario is probably uncommon, but I know several other folks do crazy things with their hardware like streaming via twitch, recording video, multi-monitor gaming, etc. In my case, I typically have the following configuration:

Monitor 1: Game running in 1080p on my center monitor (Using Starcraft 2 in my photo example)
Monitor 2: Twitch / Netflix / Amazon Instant / Youtube video streaming
Monitor 3: Chat / Gmail / Web Browser (Notebookreview of course)
Monitor 4: Benchmarking tools / HWMonitor / EVGA Precision / Task Manager Here is a pic of the workspace, and a screenshot of my typical relaxation desktop:

Spoiler

When I'm doing real work, I will have some configuration of Excel, SAS, STATA, Word, Outlook, ArcGIS, and MATLAB open at any given time. My 4 monitors are always in use, and the CPU is constantly being flogged. I will on occasion take my laptop into the field and run similar tests, thus I would prefer to have a beefy laptop rather than a desktop. For this example, the desktop comparison has an i7-4770k (3.9GHz single core turbo, 3.7GHz 4 cores.) All machines have the same other relevant specs other than CPU: 8GB DDR3-1600 9-9-9-24, 1TB Samsung EVO 840 SSD, new windows installations, and the same version of all softwares, and drivers where appropriate. All laptops are tested on a Cooler Master Notebook cooler (fan not running), all processors have been mounted with IC Diamond, both Clevos are using a 180w PSU, and set to use max fans (FN+1)

Two more things to note about my test

First: I use completion times from a statistical software package that stresses the CPU only. We’ve all seen CineBench, PCMark and other benchmark scores for various configurations, both laptop and desktop. Quite frankly, these applications don’t stress the hardware the same way Prime95 or my software does, and I don’t really care about a synthetic benchmark if it doesn’t equate to meaningful time savings in my real life. For those curious, this benchmark is creating a neural network from hundreds of thousands of observations using SAS. I do this on a near daily basis, so I want to know how much time is saved with a desktop vs a laptop, if any. This was the real reason for me running these experiments. If you want to reproduce my scenario, Prime 95 is the closest equivalent in terms of CPU load. Prime95 is NOT indicative of a typical use or a gaming scenario, but rather a max CPU power scenario with little to no GPU utilization, and 100% CPU utilization. All quoted completion times are an average of 3 runs per condition, and all runs were within 20 seconds of all other observations for each condition. It is a very consistent test that is particularly relevant for me.

Second: The relaxation usage scenario (Gaming, streaming, etc.) will work flawlessly with the desktop PC. What I mean by that is when playing Starcraft 2 or Skyrim, streaming video quality (Netflix / Twitch / ETC) is smooth, my display link monitor is responsive, and I have a constant 60FPS in the game. If I play Battlefield 4 with this scenario, my FPS will commonly drop by ~10FPS compared to gaming on just the laptop, without extra processes streaming, etc. This is at Medium Settings (the preferred settings for the W230ST which has the weakest GPU here.)

So, lets talk about the problem.I have a few machines lying around the house, and I'll show some comparisons. First, I'm using Intel XTU and Prema's bios in the P650se to try and increase the multi-core power of my i7-4710hq. In a stock configuration, this processor sports a 3.5GHz frequency for one active core, moving down to a 3.3GHz multiplier for 4 cores. The screenshot below shows XTU in a stressed, but typical usage scenario for me

Spoiler

For whatever reason, processor frequency shows the correct GHz, 0% Thermal Throttling, 0% Power limit throttling, but we have constant Current Limit throttling (Default current limit on this machine is 72A...we'll see later on my W230st this limit is only 60A.)

In my case, this leads to severe stuttering in my streaming video, my displaylink monitor lagging, very low FPS in my game (~25FPS when normally 60+), and significantly longer compute times in my work processes (24 minutes completion time compared to 19 minutes for a 4770k @ 4710hq frequencies.) The obvious solution to this is to raise the Processor current limit and power limit. With the help of premamod bios, let’s try this out--still at stock clocks. I’ll also drop my vcore by -55mv (the largest stable undervolt for my particular machine.)

Spoiler

Even with an irrational 90w specified TDP, and a 125A current limit, we still have current limit throttling of up to 15%. The processor frequency stays fixed at 3.28 as per XTU, but the stuttering is still very obvious and in similar intervals to the graph in XTU. Using throttlestop instead of XTU resulted in the same frustration. My compute time on the work process dropped to 23.5 minutes, but it is still off from the 21 min seen by the desktop processor at 4710hq frequencies. The fact that Total CPU TDP stayed below 37w the entire time is impressive for a processor binned as 47w TPD, but as we can see, it never got the current to go higher anyway. I'm not sure why my unlocked premamod is not responding to the current limit like I am used to with MQ series processors, but we'll see in the next picture, it doesn't really matter. For kicks and giggles, let’s try again with the multipliers increased.

Spoiler

In this situation, the machine suffers Thermal throttling when all 4 cores are being taxed in my work scenario. Raising the TDP, and current limit WHILE increasing the multipliers removed the previous power throttling oddly, but now we now are at the mercy of the heatsink (I'm using FN+1, pasted with IC Diamond in this case, and using a Laptop cooler).

With the P650SE, it seems there is no way to get desktop equivalent performance when I can't maintain a 45+ TDP without hitting the thermal limit—which is curious since the cooling appears better than many other machines (2 dedicated CPU heatpipes and reasonable radiator.) My frequency jumps between 3.2GHz and 3.5GHz, my compute task finishes in 23 minutes, and the desktop completes in 19 minutes (at overclocked 4710HQ frequencies of course.) More importantly, in my relaxation scenario, I am still getting stuttering in the Twitch / Netflix stream, and my FPS in Starcraft and Skyrim are nearly halved compared to the same scenario with the desktop where there was no change in gaming. It’s as if there is some sort of high-low frequency fluctuation, (similar to what nvidia cards do when they are throttling) that is leading to poor performance. In the P650se, there is always some sort of throttling identified by XTU or Throttlestop, be it current, power, or ultimately temps.

My next machine is a W230ST with an i7-4800MQ processor(3.7GHz Single Turbo / 3.5GHz quad) This is the chip that should, for all intents and purposes, accurately replicate the power of a desktop i7-4770 since it can easily overclock to match the 4770 boost clocks, but it does have only 6mb of cache compared to 8mb of cache on the 4770. The W230ST is significantly gimped with a default processor current limit of only 60 amps (this was the same regardless of a 4800MQ or 4700MQ in the machine.)

Spoiler

What's particularly interesting is with everything in a stock configuration, running my model, it begins with 100% current limit throttling with 50+ watts TDP, but after the turbo power time window expires, it changes to being TDP limited (you can see the switch in the screenshot.) Initially, it pegs the frequency to 3.39 (below the 3.5 stock advertised), and after the 28 seconds expire, it varies the frequency between 3.18 and 3.29 GHz. Unlike the 4710HQ, the 4800MQ reports 47w TDP in use for the duration of the model. My work simulation finishes even slower than the P650se at 24 minutes for the same frequencies. CPU Temps never exceed 79C, so we are well within thermal specifications. This is expected, since the W230ST shares a radiator (but not heatpipe) with a 50+ watt GPU (GTX765m.) Let’s try again but with stock clocks and power limits raised, as well as with a -45mv undervolt.

Spoiler

Unlike the P650se, when we remove the TDP limit, and the Current Limit, the CPU maintains boost clocks indefinitely with plenty of thermal headroom. Not only that, we finally the advertised clocks! CPU temp never goes above 77c (by nature of the undervolt) and the process completes in a speedy 21 minutes—much closer to the desktop, but still not equivalent.
Trying again with an overclock on the 4800MQ to match the Stock 4770k.

Spoiler

Here, we find again, no reported power throttling, no thermal barriers (always below 90c) and notice that CPU TDP is all the way up at 65w Frequencies report a constant 3.7GHz on all 4 cores. Even though frequencies are an exact match for the 4770k, the process completes in 19 minutes compared to 17 minutes, but more importantly—I am still getting micro-stutters in game, and in the fps of the steam going on in the background. BF4 FPS drop substantially (22FPS delta) and it becomes unplayable. For whatever reason, the 4800MQ still underperforms the 4770k both in benchmark test, and usability of the machine.

Trying again with a more substantial overclock on the 4800MQ

Spoiler

In order to even raise the single core multiplier up to the 4.1GHz max, we have to not only back off the undervolt, but overvolt the processor by +5mv to avoid a BSOD. The combination of higher frequencies and the overvolt lead to fluctuating 4 core frequencies between 3.39GHz and 3.6GHz, and a sudden hard power limit of 58 watts. Even though we never exceed 90c, the W230ST suddenly refused to exceed the 58w TDP limit at this voltage, and I was unable to thwart it with software (Throttlestop, Intel XTU, etc) or in the Premamod Bios. The only way to remove this spontaneous hard limit was to remove the AC adapter, battery, and pull the CMOS battery! I have no idea why this would have be triggered all of a sudden, but forcing a +5mv and max multipliers on the 4800MQ will cause it every time on my W230ST. Let’s try again with a slightly lower OC, and a slightly lower vcore.

Spoiler

Here, you can see we backed off the single core frequency to 4.0GHz, but were able to keep the 3.9GHz quad turbo. Part of this was due to resorting back to a -45mv undervolt, as well as using throttlestop to modify the Intel Power Balance, and reassign TDP to the CPU from the iGPU. This downclocks the iGPU, and provides a bit more power to the CPU. Now, normally I would say why does this matter? If the TDP is unlocked, can’t the CPU just call for more current? Well, for whatever reason, not on my machine. Without the reassignment in throttlestop, the machine crashes after modeling for 2 minutes. It seems as if the physical VRMs / motherboard is unable to provide more than 67w of power to the entire Package. When benchmarking in this scenario, the operation did not cut down on my model operation time (19 minutes.) My gaming scenario only dropped 5-10 frames on Skyrim / SC2, but the battlefield test still dropped ~18 frames compared to the desktop’s 10 (on lower frequency mind you.)

Also of note, the W230ST was pulling 95 watts from the wall—on the stock 120w PSU, there would certainly not be enough room left for the GPU’s 50w.

Spoiler

The most interesting thing here to me is that the 4800MQ had no benefit over the 4710HQ comparing the P650se and W230st. Even stock it failed to boost beyond 3.1GHz - 3.3GHz, logically because of the power requirement put in place on the stock bios. It seems shady to me that Sager would sell the 4800MQ and 4900MQ as upgrades for the W230st, when it can’t take advantage of the better processor. My guess for as to why they did this is to control heat on the shared radiator. With 65w from the processor, and it is already at 92c, the W230st cooling can’t handle the GTX765 at the same time. Perhaps the W230SS has better results. I intend on trying this with the P370sm next week when I get a hold of one with the 4800MQ, however I think we already know the answer. D2 Ultima has boosted his 4800MQ far beyond what I was able to do with the W230st. This is surely a function of the P370sm having both the cooling, and power for such a processor.

 

TL;DR:
With the desktop I get no lag in game, no stuttering in my streaming video, Netflix, or displaylink monitor when playing a game like Starcraft or Skyrim. In work scenarios, a typical job that takes a stock i7-4770k 17 minutes, will take an i7-4710hq (3.5Ghz Turbo) 23 minutes and my overclocked 4800mq (4 cores @ 3.9Ghz) 19 minutes. Even with the 4800MQ at the same clocks as the desktop i7-4770, it is ~18% slower to complete the task. I'll talk about a few potential reasons for this in the next post, but I want to hear if others have ideas why the desktop chip is still out performing the mobiles despite the same frequency. 18% Is significant to me, but I admit it is only in very specialized situations.

 

Here are some of my thoughts on why this maybe be happening:

- Power limits we aren’t able to see, either immediate requests for a lot of current the motherboard is unable to supply, or sustain?

- CPU Cache (8mb vs 6mb)

- iGPU contributing towards TDP (Optimus on Clevos vs. dedicated on Desktop

- Individual hardware limits set on a per machine basis (P370sm vs P150sm vs W230ST, etc)

 

The 4810MQ has a TDP of 47W where as the desktop 4770 has a TDP of 84W, the 4810MQ was never going to turbo like the desktop chip to be fair. At stock the chip had a short 57W window and then moved to 47W which is exactly what it should do. Also I believe they are partially TDP unlocked and allow going from 47W/57W to 57W/67W.

The advantage of the desktop K processor is the full ability to tune all of these values while not paying for a mobile extreme It really is designed for people like you who need the horsepower.

 

Lots of information and time taken here, thank you for doing so!

 

You're obviously right regarding Intel's designed TDP and turbo behavior for these processors. It is exactly what intel designed them to do from a stock perspective.

My point here was two fold: to show that a mobile CPU != desktop CPU just because they have the same frequency. This idea has been passed around in the recent P750zm thread...it's not just frequency, its the power the chip can consume and dissipate). Second, I wanted to show that you can't compare a 4800MQ in one machine and expect it to do the exact same thing in another--even at stock frequencies (as was the case in the W230st.) I know the 4800MQ can do a lot more than what it does in my W230st.

This is especially true when we get outside the realm of factory settings on processors, the max TDP allotted via power delivery and cooling capacity by the specific laptop seems to be more relevant than the specific chip when it comes to heavy multi-threaded applications. What really makes the difference between a 4930MX and a 4800MQ when both are consuming 65-75w?

I'm fairly certain D2 has a 4800MQ consuming up to 80w TDP (obviously not stock.) My real curiosity here is whether a 4800MQ sucking down 80w in a P370sm will out perform something like a i7-4790s in the P750zm sucking down 65w, and locked to that 65w.

Also, what do you mean by "they" are partially TDP unlocked? Are you referring to the processors? It seems some of them (the 4800MQ at least) are completely unlocked with regards to TDP, while others (4710HQ) are locked down pretty tight...either in the CPU itself, or the motherboards designed specifically around the soldered CPU.

Thanks for the comments; I really appreciate it.

 

Ok well the older 3 and below series CPUs were TDP limited to +10W regardless of what the slider says, I honestly have not tested to the 4xxx series for this behaviour, I would be interested in seeing a screenshot showing the TDP of the 4800MQ at 80W, if this is the case then intel have changed the behaviour of a least some of the chips.

 

PLEASE use spoiler tags to hide those pictures. It's a huge scrolling post and it doesn't need to be.

Power limits are relaxable. I set mine to draw 100W max and it has been known to draw far over 67W. The 67W limit you probably are thinking of is some machines' limits in what the board will supply to the CPU. Example is my 4800MQ with TDP limits set to 100W pulling 80W in linpack at stock speeds (3.5GHz) with a -50mV undervolt:

Spoiler

Also, to go further to that, if I set voltage to "default" and OC it to 3.8GHz and run the same bench, it can pull as far as 94W:

Spoiler

1 - Better question is 4900MQ versus 4930MX. 4800MQ has 6MB L3 cache, and thus may very well underperform in certain synthetic benchmarks which rely on cache for something. In applications that DON'T rely on that, such as x264 rendering, they should be the same.

2 - Yes, I have the pictures above this direct quote. Also, I have proof that my 3.8GHz 4800MQ outperforms the 4790S in 3Dmark Firestrike's Physics test. For reference (look at the Physics scores):
Bigspin's 4790S benchmark
My 4800MQ benchmark (3.8GHz)

3 - Some motherboards lock TDP down for the CPU, as I mentioned above. The M17xR4 (I think that's the machine) won't allow more than 67W to be fed to the CPU, as far as I remember. So even with an extreme CPU, you're still limited. But the chips themselves should not be. The HQ chips however... those may very well be partially TDP locked. The 4710HQ chips have been known to not hit their 3.5GHz max 4-core turbo when OC'd. The other HQ chips may have similar downsides, but I do not know for certain. If anybody with a GT72 or GT80 can do some testing for us with the 4980HQ, we'll find out for sure. But then again, since HQ and MQ can't fit in the same board, we have no way of knowing if things are board limited or what. I am much more willing to bet that it is the fault of the HQ chips however, due to their existence in many lower-end consumer machines like HP/Dell/Acer/etc lines, which are designed to be locked down. Having the chip be partially locked would indeed help stave off people who are keen at getting around locks.

 

Slight OT; have you managed to get ArcGIS Pro running and, if so, do the main tools deploy MT now? Use 10.3 on the W230SS for now: :zzz: ... and the only upgrades are of no use.

 

I use a np9377sm-a with i7 4910mq. I run prema mod bios and run my ram at 2133 CL11. I have had zero issues with power or current throttling with my cpu overclocked up to 4.2G. Temps are another problem though. So far I only tried benching with XTU. I score 1083 at 4.2G.
For 24/7 usage I clock in at 4G on all 4 cores loaded so I can maintain temps under throttling with fans on full for all my apps. I can adjust power and current limits to whatever is needed under XTU benching, so far that's 110W and 140 amps. There's no throttling for the duration of the benchmark.
Can you run XTU benchmark for the 4770k at 4.2G and report back with the score as I'm curious on the comparison given both cpus have 8MB cache.

 

Be wary that XTU is VERY memory dependent. CPU speed actually has quite little to do with it. If you look at the top benches for the 4800MQ, they're *ALL* below 3.5GHz max speed... but they have CL6 dual channel RAM. Your CL11 2133MHz is doing wonders for your score.

 

I haven't tried pro. I just updated to December 10.3 on my P650se. I tried 10.1 on the W230ST, but deregistered it pretty quick because of some frustrations with display scaling on the tiny (for my eyes) 1080p display.

 

That's probably true as each multi between 4G & 4.2G was only giving me an additional 20 marks or so. Still I see someone with a 4940MX scoring higher than me with a clockspeed of 4,368 amd memory speed of 2,215 only 14 points higher though.

It does show that some Clevo models are capable of sustaining much higher TDP/power/current when overclocked.

 

Thanks, will hope for a CS5>CS6-style improvement upgrade then. Or, perhaps, may have to put some effort in getting Pro to run ... but it's too beta right now.

 

This is why you have to be careful focusing on a single benchmark.

 

The screenshots you wanted are on the previous page

 

Thanks I'll take a look