M1 Air with Thermal Pad Mod

[KnICS]

Great results juicy.

For a very quick ad-hoc test of my franken-cooler, I ran Cinebench r23 while on a Teams call (I have to work for pay for these crazy hobbies), with some office apps, preview, safari and chrome open. The performance after 10 minutes was 7155 although with all that other stuff going on we cant trust that score. However, I did this to drive CPU heat up as high as I could get it.

It topped out at 98C and two cores did throttle, but the battery reading didn't go above 31C (ambient here is 29). This indicates that the high battery temps are bleeding from CPU to the battery and not caused by excess battery use. With the thermal pad in place, that heat goes tot he case first since it conducts better than air, and is then cooled. With the cooler off, it will bleed via that case and also internal air at different rates, which is why battery heat rises regardless of the thermal pad, just at different rates.

For consistency with my previous tests, I'll be forced to play DOS II for another 45 minutes but not until after work. I can test R23 over time with the cooler running as well.

Final components will be in next week, and at that point I'll post pics

 

[Leon1das]

Update from a less brave M1 MBP and MBA owner.

I did not dare to mod MBA yet - but I stress-tested MBP via Rosetta gaming (Bioshock Remastered) with adding TGPro for fan control

System default fan control in MBP is very "permissive" - and will not kick in timely at all.
Thats why most of the users say that MBP is silent even with having the fan.

But then I set-up gaming session of 35min with 2 profiles:
- first part with adjusted more aggressive cooling
- second part with 100% fan (7200rpm)

You can see the result in TGPro log which i converted to excel table, and marked (in orange) the moment when aggressive fan kicked in.
Only with aggressive fan and full stress - battery settles at 36 C, and SSD is 44-45 C.

 

[Mike Boreham]

If thermal pad mod with MBA gives better results for battery too - it may make sense to do it with MBP too...

My reading of previous posts is that thermal pad mod tends to increase battery temp.

 

[Leon1das]

My reading of previous posts is that thermal pad mod tends to increase battery temp.

Hi I just edit the post with adding the TGPro log and sharing slightly updated results.
Sorry for this.

 

[KnICS]

I've been stress testing my M1 Air with the thermal pad mod on, and off, and on again. Battery temperatures don't seem to get any hotter than without the mod, although depending upon the circumstances it might reach the higher temps sooner. I'm going to keep experimenting and will post results here for anyone who wants to know.

BTW, I have no desire to convince anyone to do this mod. Some people like to mod, some don't. Personally, I want to learn as much as possible about the thermal headroom with the new M1 designs because what we learn now will likely have even greater impact in future MacBook models that use more power.

For the current Air, the mod provides little value in terms of performance because well ... the performance is already amazing, as is the multi-tasking response. An extra 5-10% is just icing on the cake. The mod will preserve the CPU if we can keep the temps lower even by a few degrees. You'd assume Apple would have the chip throttle with the correct tolerances already, but I've seen the CPU temp jump as high as 102C and that just doesn't seem right.

@Leon1das, testing load with games is much more fun than running Cinebench, isn't it? Bioshock is actually fairly well tuned even though its running through Rosetta2. So far, Divinity Original Sin 2 is the most demanding game I've found. According to Activity Monitor it will actually use 150%+ of total CPU and 100% GPU and it never seems to back down.

 

[denzdaniel]

Hey guys I uploaded a video of my M1 MacBook Air Thermal Pad MOD and here it is...
See the Before and after! You'll be amazed!😉

This is just my experiment and I'm just sharing it with you.✌🏼

 

[Tenkaykev]

Hey guys I uploaded a video of my M1 MacBook Air Thermal Pad MOD and here it is...
See the Before and after! You'll be amazed!😉

This is just my experiment and I'm just sharing it with you.✌🏼

Hats off for doing this. A couple of observations / queries. I wouldn't be happy removing the cover of the heatsink as you did, the piece of black tape seems to have metal foil attached and I'm not sure there'd be much to gain by removing it as opposed to just putting the thermal pad on top of it. As the heat from the processor is travelling into the heatsink plate and travelling to the square part where it makes contact with the rear of the case via a thermal sponge, I wonder what the effect would be of replacing the thermal sponge with a thermal pad.

Great work though, many thanks for the video 👍

 

[timmillea]

Thanks to all who contributed to this discussion. Having discovered it today and read through it all, going off for background here and there, it would appear the most bang per effort and least risk of affecting the warranty is to simply add a 2mm thermal pad on the 'higher' main CPU level of the heatsink. Taking measurements from the iFit photos, it appears that main section of the heatsink requires a thermal pad 100mm x 39mm with a corner cut out.

The hotter bottom case also does not bother me because 1) when under sustained load, it is going to be plugged into power on a desk and probably left all night unattended and 2) it indicates that the mod works - i.e. the lower case is working as a better heatsink! I would hardly be running heavy loads on it with it on my lap - or running it on battery power. I take care of my batteries!

In terms of thermal transmission air has a value of approximately 0.026 W/m.K, so even cheap thermal pads will be of the order of 100 times better this (at around 2.5W/m.K). With the M1 having such a low TDP, a cheap thermal pad such more than suffice because 2.5W/m.K with a gap of 2mm yields 1250W/K (vs. just 13W/K with air).

It would also appear to make sense for sustained loads to use the MBA in clamshell mode, upside side down to better dissipate the heat spread to the lower case.

I've ordered my thermal pad from the bay. Does anyone recommend a free M1-compatible temperature monitor for testing? My trustee Macs Fan Control - with readout of every Mac temperature sensor that kept my 2008 MP going as long as it did - appears not to be compatible.

My main interest is Handbrake Blu-ray to H.265 (software) conversions, which appear to throttle back approaching 40% - but I am happy to run the CineBench R23 benchmark scores for comparison and share them here if there is any interest. I have the base model M1 MBA (8GB/256GB 7-core GPU) manufactured 1st week of November. I will also find out if my base model has Apple's pre-installed silicone foam pad attached to the bottom case.

 

[Tenkaykev]

Thanks to all who contributed to this discussion. Having discovered it today and read through it all, going off for background here and there, it would appear the most bang per effort and least risk of affecting the warranty is to simply add a 2mm thermal pad on the 'higher' main CPU level of the heatsink. Taking measurements from the iFit photos, it appears that main section of the heatsink requires a thermal pad 100mm x 39mm with a corner cut out.

The hotter bottom case also does not bother me because 1) when under sustained load, it is going to be plugged into power on a desk and probably left all night unattended and 2) it indicates that the mod works - i.e. the lower case is working as a better heatsink! I would hardly be running heavy loads on it with it on my lap - or running it on battery power. I take care of my batteries!

In terms of thermal transmission air has a value of approximately 0.026 W/m.K, so even cheap thermal pads will be of the order of 100 times better this (at around 2.5W/m.K). With the M1 having such a low TDP, a cheap thermal pad such more than suffice because 2.5W/m.K with a gap of 2mm yields 1250W/K (vs. just 13W/K with air).

It would also appear to make sense for sustained loads to use the MBA in clamshell mode, upside side down to better dissipate the heat spread to the lower case.

I've ordered my thermal pad from the bay. Does anyone recommend a free M1-compatible temperature monitor for testing? My trustee Macs Fan Control - with readout of every Mac temperature sensor that kept my 2008 MP going as long as it did - appears not to be compatible.

My main interest is Handbrake Blu-ray to H.265 (software) conversions, which appear to throttle back approaching 40% - but I am happy to run the CineBench R23 benchmark scores for comparison and share them here if there is any interest. I have the base model M1 MBA (8GB/256GB 7-core GPU) manufactured 1st week of November. I will also find out if my base model has Apple's pre-installed silicone foam pad attached to the bottom case.

That seems very logical. I have exactly the same setup base M1 Air. I wondered what your thoughts are about simply replacing the foam pad with a more thermally efficient material first, getting some benchmarks and then comparing with the mod you propose?
Looking forward to the results 👍

 

[timmillea]

That seems very logical. I have exactly the same setup base M1 Air. I wondered what your thoughts are about simply replacing the foam pad with a more thermally efficient material first, getting some benchmarks and then comparing with the mod you propose?

I don't want to do anything inside that might be detectable should I ever need a repair under warranty. I have a two-year warranty which is too valuable to me to risk so this also means minimising opening and closing the back more than I have to, i.e. hopefully just the once unless it needs repair, in which just the twice . If there is a piece of Apple-installed thermal foam on this model, I can imagine it won't come off cleanly. Besides, running it in 'upside down clamshell mode' last night and feeling the case with my hands (totally unscientific) by far the hottest area was exactly above the M1 - about 1" from the hinge and around half an inch offset towards the sound port - the heat is currently that localised. I would imagine, if there is no foam there, adding a 3mm thermal pad would make a near insignificant difference compared to a pad directly above the M1.

Also I notice the purpose of that thicker half of the heatsink not over the logic board appears to be to transfer heat to the more substantial ridge of the upper case (between the keyboard and screen). This gets far hotter than the lower case under heavy loads (without the mod), either way up, and with the heat more spread out.

My thermal pad should arrive by Monday. I will update my test results and observations early next week.

 

[Tenkaykev]

I don't want to do anything inside that might be detectable should I ever need a repair under warranty. I have a two-year warranty which is too valuable to me to risk so this also means minimising opening and closing the back more than I have to, i.e. hopefully just the once unless it needs repair, in which just the twice . If there is a piece of Apple-installed thermal foam on this model, I can imagine it won't come off cleanly. Besides, running it in 'upside down clamshell mode' last night and feeling the case with my hands (totally unscientific) by far the hottest area was exactly above the M1 - about 1" from the hinge and around half an inch offset towards the sound port - the heat is currently that localised. I would imagine, if there is no foam there, adding a 3mm thermal pad would make a near insignificant difference compared to a pad directly above the M1.

Also I notice the purpose of that thicker half of the heatsink not over the logic board appears to be to transfer heat to the more substantial ridge of the upper case (between the keyboard and screen). This gets far hotter than the lower case under heavy loads (without the mod), either way up, and with the heat more spread out.

My thermal pad should arrive by Monday. I will update my test results and observations early next week.

Completely understand wanting to do a reversible mod. Thanks for doing this, I'm eager to see how things pan out.👍

 

[Mike Boreham]

Does anyone recommend a free M1-compatible temperature monitor for testing?

I am happy to run the CineBench R23 benchmark scores for comparison and share them here if there is any interest.

Not aware of a free one but TGPro is not expensive. I use iStat menus, more expensive.

I have done 3 x 10 minute continuous Cinebench runs in quick succession on my M1 MBA:
First run: 7492
10 mins: 7020 (-6%)
20 mins: 6774 (-10%)
30 mins: 6632 (-11%)

These were run with machine "as is", ie all background process left running.

 

[Tenkaykev]

Not aware of a free one but TGPro is not expensive. I use iStat menus, more expensive.

I have done 3 x 10 minute continuous Cinebench runs in quick succession on my M1 MBA:
First run: 7492
10 mins: 7020 (-6%)
20 mins: 6774 (-10%)
30 mins: 6632 (-11%)

These were run with machine "as is", ie all background process left running.

It looks as though it throttles after about 10-15 minutes and then stabilises.

 

[Mike Boreham]

Not aware of a free one but TGPro is not expensive. I use iStat menus, more expensive.

I have done 3 x 10 minute continuous Cinebench runs in quick succession on my M1 MBA:
First run: 7492
10 mins: 7020 (-6%)
20 mins: 6774 (-10%)
30 mins: 6632 (-11%)

These were run with machine "as is", ie all background process left running.

Repeated this as a single 30 minute run. CB carries on starting runs within 30 minutes then finishes the last run it started, so total time was actually 32 minutes.

Very close to previous results. Not quite flat after 32 minutes. Note this graph does not have a true (zero) origin so makes throttling look much more than it is....only 10% after 30 mins with all eight cores on 100%.

The "pACC MTR 3" sensor fluctuated between 95C and 98C from about the third run onwards.

I guess Handbrake might load the M1 chip more than Cinebench.

Not planning on doing any throttling mods.

 

[denzdaniel]

Not aware of a free one but TGPro is not expensive. I use iStat menus, more expensive.

I have done 3 x 10 minute continuous Cinebench runs in quick succession on my M1 MBA:
First run: 7492
10 mins: 7020 (-6%)
20 mins: 6774 (-10%)
30 mins: 6632 (-11%)

These were run with machine "as is", ie all background process left running.

Hi Did you do the MOD? How come the score went down? or decreasing each run?

 

[Mike Boreham]

Hi Did you do the MOD? How come the score went down? or decreasing each run?

No not done the mod.

Score goes down as machine warms up causing more throttling.

In case you misunderstood, these were not three tests at 10 minute intervals. They were three consecutive 10 minute duration tests. The machine was running continuously with all cores near 100%. The scores are what was being achieved at 10, 20 and 30 minutes.

 

[denzdaniel]

No not done the mod.

Score goes down as machine warms up causing more throttling.

In case you misunderstood, these were not three tests at 10 minute intervals. They were three consecutive 10 minute duration tests. The machine was running continuously with all cores near 100%. The scores are what was being achieved at 10, 20 and 30 minutes.

Thanks for clarifying....

 

[timmillea]

A small update. I have established my testing regimen. I had never used CineBench before so I had to play around with it, find out what it did and what its results mean. So I have conducted the 'before mod' tests today. After a few trial runs, I have chosen to run a custom minimum 60 minute duration and to record temperatures of the CPU and battery for 20 mins after CineBench has finished. A few observations so far.

1) CineBench averages out the performance per pass. My experience with Handbrake shows the M1 throttles back drastically (at least 40%) after 10-20s seconds of high load. CineBench's around 2 minutes per pass on the M1 all but loses that initial cliff in its first pass score.

2) I extended the minimum testing time from the regular 30 minutes to 60 minutes because there were still steady incremental decreases in CineBench scores at every completed pass. I really wanted to find the 'steady state' performance if left on maximum load indefinitely.

3) CB scores however are average per pass from the start of the test - meaning that scores will forever trend towards steady state but never reach it as the initial fast pace is always incorporated into the average. I attach a chart showing the difference between the scores each pass as reported by CB (i.e. averaged from the the start), compared to calculating the CB score just for the most recent pass. You can see the average (blue line) is trending towards the last pass score (green). The latter however gives a better indication of performance at any point in time, under sustained maximum load.

4) As the CPU reaches max temparature (around 93.5 degrees) within seconds and then stays within a tight limit for the remainder of the hour-long test, the throttling back, which occurs over a much longer timeframe, appears to be controlled by the other temperature sensors, e.g. battery temperature (maxes out and maintained around 41 degrees C). On another rough and ready chart, also attached, the battery temperature in green (x150 to be on the same scale) is almost the mirror image of per-pass performance.

I will have my 'post mod data' from tests under identical conditions just as soon as my pentalobe P5 screwdriver arrives.

My thoughts are that if the mod merely conducts the M1's heat better to the lower case but the MacBook throttles down according to battery temperature, then the mod may just buy an extra minute or so of short-term 'maximum load' performance but its steady state, maximum load performance will be determined by how cool the bottom case can be kept. As is, once the heat reaches the battery, those longer than an hour CineBench figures will all trend together, one way or another.

As I say, my thoughts on studying my pre-mod data today.

Tim.

P.S. final results will have labels on the charts - please read the text for these quick and dirty ones.

 

[Tenkaykev]

A small update. I have established my testing regimen. I had never used CineBench before so I had to play around with it, find out what it did and what its results mean. So I have conducted the 'before mod' tests today. After a few trial runs, I have chosen to run a custom minimum 60 minute duration and to record temperatures of the CPU and battery for 20 mins after CineBench has finished. A few observations so far.

1) CineBench averages out the performance per pass. My experience with Handbrake shows the M1 throttles back drastically (at least 40%) after 10-20s seconds of high load. CineBench's around 2 minutes per pass on the M1 all but loses that initial cliff in its first pass score.

2) I extended the minimum testing time from the regular 30 minutes to 60 minutes because there were still steady incremental decreases in CineBench scores at every completed pass. I really wanted to find the 'steady state' performance if left on maximum load indefinitely.

3) CB scores however are average per pass from the start of the test - meaning that scores will forever trend towards steady state but never reach it as the initial fast pace is always incorporated into the average. I attach a chart showing the difference between the scores each pass as reported by CB (i.e. averaged from the the start), compared to calculating the CB score just for the most recent pass. You can see the average (blue line) is trending towards the last pass score (green). The latter however gives a better indication of performance at any point in time, under sustained maximum load.

4) As the CPU reaches max temparature (around 93.5 degrees) within seconds and then stays within a tight limit for the remainder of the hour-long test, the throttling back, which occurs over a much longer timeframe, appears to be controlled by the other temperature sensors, e.g. battery temperature (maxes out and maintained around 41 degrees C). On another rough and ready chart, also attached, the battery temperature in green (x150 to be on the same scale) is almost the mirror image of per-pass performance.

I will have my 'post mod data' from tests under identical conditions just as soon as my pentalobe P5 screwdriver arrives.

My thoughts are that if the mod merely conducts the M1's heat better to the lower case but the MacBook throttles down according to battery temperature, then the mod may just buy an extra minute or so of short-term 'maximum load' performance but its steady state, maximum load performance will be determined by how cool the bottom case can be kept. As is, once the heat reaches the battery, those longer than an hour CineBench figures will all trend together, one way or another.

As I say, my thoughts on studying my pre-mod data today.

Tim.

P.S. final results will have labels on the charts - please read the text for these quick and dirty ones.

This is stellar work, thank you so much for doing such a comprehensive test and sharing your methodology and reasoning. My modified glass bathroom scale with cooling fan kludge ( referred to up thread ) blows air against the bottom of the case and serendipitously might be a better way of keeping the batteries cool. I'm looking forward to the next stage of your testing. 👍

 

[Mike Boreham]

A small update. I have established my testing regimen. I had never used CineBench before so I had to play around with it, find out what it did and what its results mean. So I have conducted the 'before mod' tests today. After a few trial runs, I have chosen to run a custom minimum 60 minute duration and to record temperatures of the CPU and battery for 20 mins after CineBench has finished. A few observations so far.

1) CineBench averages out the performance per pass. My experience with Handbrake shows the M1 throttles back drastically (at least 40%) after 10-20s seconds of high load. CineBench's around 2 minutes per pass on the M1 all but loses that initial cliff in its first pass score.

2) I extended the minimum testing time from the regular 30 minutes to 60 minutes because there were still steady incremental decreases in CineBench scores at every completed pass. I really wanted to find the 'steady state' performance if left on maximum load indefinitely.

3) CB scores however are average per pass from the start of the test - meaning that scores will forever trend towards steady state but never reach it as the initial fast pace is always incorporated into the average. I attach a chart showing the difference between the scores each pass as reported by CB (i.e. averaged from the the start), compared to calculating the CB score just for the most recent pass. You can see the average (blue line) is trending towards the last pass score (green). The latter however gives a better indication of performance at any point in time, under sustained maximum load.

4) As the CPU reaches max temparature (around 93.5 degrees) within seconds and then stays within a tight limit for the remainder of the hour-long test, the throttling back, which occurs over a much longer timeframe, appears to be controlled by the other temperature sensors, e.g. battery temperature (maxes out and maintained around 41 degrees C). On another rough and ready chart, also attached, the battery temperature in green (x150 to be on the same scale) is almost the mirror image of per-pass performance.

I will have my 'post mod data' from tests under identical conditions just as soon as my pentalobe P5 screwdriver arrives.

My thoughts are that if the mod merely conducts the M1's heat better to the lower case but the MacBook throttles down according to battery temperature, then the mod may just buy an extra minute or so of short-term 'maximum load' performance but its steady state, maximum load performance will be determined by how cool the bottom case can be kept. As is, once the heat reaches the battery, those longer than an hour CineBench figures will all trend together, one way or another.

As I say, my thoughts on studying my pre-mod data today.

Tim.

P.S. final results will have labels on the charts - please read the text for these quick and dirty ones.

You are obviously being very thorough, but some things you say about Cinebench are surprising to me.

When you say "CineBench averages out the performance per pass" and "CB scores however are average per pass from the start of the test" what are you referring to as the CB reported result? When I did the 30 min test in my post#89 above I used the result of each 2 minute run, as shown in the dropdown that appears if you hover the cursor over your computer name on the left side. As in attached screenshot. Is this what you were looking at? You have to re-hover after each run to refresh it to get the last run. At the end of the 30 minute test the number that is shown where the word "Running" appears is the one for the last run.

Not sure this will affect your hypothesis.

 

[timmillea]

You are obviously being very thorough, but some things you say about Cinebench are surprising to me.

When you say "CineBench averages out the performance per pass" and "CB scores however are average per pass from the start of the test" what are you referring to as the CB reported result? When I did the 30 min test in my post#89 above I used the result of each 2 minute run, as shown in the dropdown that appears if you hover the cursor over your computer name on the left side. As in attached screenshot. Is this what you were looking at? You have to re-hover after each run to refresh it to get the last run. At the end of the 30 minute test the number that is shown where the word "Running" appears is the one for the last run.

Not sure this will affect your hypothesis.

Yes, the approximate 2-minute score updates are the average per pass to that point. They do not tell you the score of that particular pass, it has to be calculated:

(latest score x total duration so far - previous score x total duration to its scoring) / duration of last pass.

If you set a custom (minimum) duration of 10 hours or so then you could be reasonably confident that the score reported by CB is very close to, but still slightly above, the 'steady state' score. I wanted to do it faster than that and get accurate per pass results to make a pretty chart and better understand the performance with time cf. other indicators. I suspect the mod will have negligible difference over a long intensive compute, e.g. a stack of Handbrake compressions running overnight, without the intervention of cooling the bottom case. Tenkaykev's fan could be the answer for such scenarios.

Now I have set myself up nicely to be embarrassed by the actual mod data - to follow soon....

 

[Mike Boreham]

Yes, the approximate 2-minute score updates are the average per pass to that point. They do not tell you the score of that particular pass, it has to be calculated:

(latest score x total duration so far - previous score x total duration to its scoring) / duration of last pass.

Since your previous post I have searched for anything on this subject, without success. I have seen nothing which says that, for a 10 min/30min test the result is the average of previous passes. Can you post any links that have led you to this?

Can you confirm that you are looking at the numbers that appear in the hover > dropdown as in my screenshot? Why do you say these are approximate?

 

[timmillea]

2) Yes, I don't just hover but click to ensure I get the latest score from the pass just completed.
1) No links, no references. It is obvious from the numbers. The 'final score' is clearly intended to be a measure of performance over the full duration of the test. All the intermediate scores clearly lead there.

The makers of CineBench have decided to allow different duration tests. To make tests of different durations comparable the score is averaged per pass.

Try a 30 minute test using my formula above to calculate the score per individual pass and then compare the last one with the final result of a 10 hour CB test. They will be within 0.2% of each other.

You may reference this thread in future

 

[Mike Boreham]

Thanks, but afraid I am not convinced. To me it is counter intuitive and not obvious from the numbers at all.

This may not be what you intended but I just tried applying your formula (latest score x total duration so far - previous score x total duration to its scoring) / duration of last pass .... to my 30 min test data in post#89.

I used the data from the second and third data points (2 mins and 4 mins) ie early on in the 30 minute run as this is where the rate of change is greatest so any effects of averaging would be more obvious.

Score after 4 mins: 7358
Previous score (after 2 mins): 7537
Duarationof last pass: 2 mins

(7358 x4) - (7537x2) / 2 = 7179

Not seeing how this proves that the CB score is the cumulative average.

I am probably misunderstanding your sentence "Try a 30 minute test using my formula above to calculate the score per individual pass and then compare the last one with the final result of a 10 hour CB test. They will be within 0.2% of each other." I haven't done a 10 hour CB test (and don't intend to!) but I can well understand that the curve is very flat after 10 hrs and differences with the 30 minute test small, but don't see how this relates to the question of whether CB is reporting cumulative averages for multi pass tests.

It would be interesting to hear whether other people think CB reports averages, maybe I am alone in this and it is obvious to everyone else!

 

[timmillea]

While we are waiting for others to participate...

Those figures, apart from the rounding to 2 minutes, sound reasonable. They indicate, as one would expect, that the actual performance on the last pass is inevitably less than the average performance over the whole test. The average never gets as low as the current performance, well not short of a very long test and even then only because of rounding error of the CB scores.

However, I was precise with my timings, taking them from the CineBench 'count-down clock' exactly, to the second, each time a pass was completed. As I indicated, I had a fair few dry-runs before my skills were up to speed.

If you repeat the test with precise timings you will find that (the first pass score + the calculated 2nd pass score according to my formula) divided by 2 = the CB given score after the second pass, and so on.