SMC sensor keys for Apple M3 and M3 Pro/Max

[tycho]

Describe the bug
I am sure this will be a fun one to figure out, but it would be nice if Stats could recognize and handle the SMC sensors for the Apple M3 (and Pro/Max) chips. The keys are definitely different from M1 and M2. Here's what Stats currently sees (during a CPU-intensive build, so the sensors look a bit alarming):

It's pretty clear it's not reading the right keys, particularly for CPU P-core/E-core temperature sensors and the GPU.

I've written a little script to try and identify the right sensors, but there is a lot of data making it difficult to correlate. Some of the SMC values are likely not instantaneous values but rather aggregates over certain time windows (e.g. moving averages or similar).

Here's the work-in-progress script. I've added some comments to it to explain the design. I also have a few guesses in there for what some of the keys correlate to.

https://gist.github.com/tycho/73fb64e7ba0b259923a92a4bacea9e78

If you have any tips on identifying the right sensors for temperature, power, etc, I'd appreciate suggestions.

Details:

• Device: MacBook Pro late 2023, Apple M3 Max
• macOS: 14.1.1
• Application version: 2.9.12

[exelban]

Hi.
Yeap, looks like each new SoC from Apple has different sensor keys.
You can monitor all keys in the Stats by enabling the Show unknown sensors option in the settings.

If you are sure the keys you have in the script is the correct I can add them to the Stats.

[tycho]

The ones I have in the script are currently just guesses. I am somewhat confident the P-core/E-core temperature ones are right, but I'll keep tinkering with this to build some more confidence. I'm pretty sure the GPU-related ones are wrong, I'm still working on hunting those down.

[exelban]

Thanks. I will add CPU cores to the Stats. Please let me know if you will find more sensors.
Unfortunately there is a way to get sensor keys: guessing(

[exelban]

PS: you have duplicated Tf0A and Tf4A keys

[tycho]

PS: you have duplicated Tf0A and Tf4A keys

Indeed I did, whoops. The E-core sensors are hard to figure out because of how challenging it is to get a heavy enough user workload to stay running on those (even with "background" QoS class, low niceness, etc). Maybe I'll try triggering a Spotlight rebuild or something, that should put some load on them. And I know that is pretty sticky to E-cores.

[tycho]

Okay, I updated the script and ran some more workloads to try and figure out the E-core/P-core/GPU temperature sensors.

The P-core ones I'm pretty certain are right.

The E-core ones are probably correct, but it's hard to tell. Even under load (Spotlight reindexing and Time Machine backup), they only get as high as about 54C and idle at 48C. I didn't see any other sensors that were obviously correct though.

As for the GPU ones, they're very definitely correlated (rising and dropping based on the GPU workload), but there sure are a lot of them. I mean, yes, there are 40 GPU cores in this thing, but 14 sensors for that seems like too many. EDIT: Pared the GPU sensors down to just eight -- six of them were duplicating values from other keys.

As for power sensors, I think I might have a lead on the GPU power value.

Comparing with output of powermetrics, watching it in lockstep with watch-smc.py, I suspect that GPU total power is a sum of the PCxx sensors:

PC00:   0.000   L  +0.000 B  +0.000
PC01:   0.000   L  +0.000 B  +0.000
PC02:   0.000   L  +0.000 B  -0.002
PC03:   0.000   L  +0.000 B  +0.000
PC10:   6.200   L  +0.507 B  +6.200
PC11:   0.000   L  +0.000 B  -0.001
PC12:   6.373   L  +0.592 B  +6.373
PC13:   0.000   L  +0.000 B  +0.000
PC20:   5.930   L  +0.229 B  +5.930
PC21:   0.000   L  -0.000 B  +0.000
PC22:   5.870   L  +0.342 B  +5.870
PC23:   0.000   L  +0.000 B  +0.000
PC30:   0.000   L  +0.000 B  +0.000
PC31:   0.000   L  +0.000 B  +0.000
PC32:   5.254   L  -0.567 B  +5.161
PC33:   0.000   L  +0.000 B  +0.000
PC40:   7.068   L  -0.099 B  +7.068
PC41:   0.000   L  +0.000 B  +0.000
PC42:   0.735   L  -0.167 B  +0.607
PC43:   0.065   L  -0.002 B  +0.064
PCMC:   0.000   L  +0.000 B  +0.000
PDTR:  83.598   L  +0.028 B +65.862
PHPC:  58.807   L -16.128 B +54.133
PHPS:  52.186   L  -0.050 B +49.534
PMVC:  17.747   L  -0.473 B +12.691
PPSC:   8.786   L  -0.479 B  +7.822
PR2b:   1.708   L  -2.956 B  +1.673
PSTR:  89.598   L  +1.378 B +72.291
PSVR:  58.627   L  +1.125 B +56.845
PZC0:  52.098   L  -0.165 B +49.449
PZC1:  52.723   L  +0.982 B +50.117
PZD1:   8.261   L  +0.186 B  +7.812

The above is when powermetrics claimed GPU Power: 30684 mW. The sum of the PCxx sensors is 37.495. A bit higher than what powermetrics is reporting, but if you subtract out the sum of PC4x, you get a total of 29.627, which is really close. Based on issues I've seen with other projects, such as asitop, I think powermetrics is not being well-maintained by Apple. It's possible that it summed up PC0x through PC3x and incorrectly excluded PC4x from its total.

[exelban]

@tycho it's a pretty nice script)
Thanks for sharing)

Will try to find an m2 platform to get sensors for that soc. Because m2 has a pretty big gap in the sensors.

I want to add this functionality to the app now😅

[tycho]

I've got an M2 MacBook Air I could tinker with getting some more sensor info from. I'll have to see what I can find.

[gilby101]

On my M3 (not Pro or Max) my guesses using smc (and comparison with watch-smc and values.swift) are:

P-Cores: There are no Tf keys - so completely different. I think the 4 cores are:
Tp01 Tp05 Tp0D Tp0H which repeat very closely as Tp0L Tp0P Tp0X Tp0b
This is more like your base Apple silicon keys.

E-Cores: I am happy with Te05 and Te0L (as in your list). I think Te0P is wrong and Te0S does not exist.
I don't know where the other 2 E-cores might be!

[tycho]

Interesting! Could you share the output of "/Applications/Stats.app/Contents/Resources/smc list"? It would be interesting to know which keys the M3 and M3 Max have in common -- any added keys in the M3 Max might be the additional cores and could give hints as to the naming scheme.