Nuvia: don’t hold your breath

[Artemis]

Could you elaborate? I'd be interested to know more. It's hard to empirically check the numbers coming from powermetrics as there's no other way to measure power (other than wall power) of a Mac

Yes, the wall or someone testing the actual VRMs/power delivery is the best way to do this. That’s what Anandtech did for years with measuring A series power dominating rivals, you measure everything on down from the SoC to the package and DRAM to VRM/power stuff. These are all part of the product in a meaningful way, like how power costly DRAM accesses will vary based on L2 and SLC cache amounts, stuff like that.

Apple's M1 Pro, M1 Max SoCs Investigated: New Performance and Efficiency Heights

www.anandtech.com

And well (see power section) powermetrics is modeled internally and it ain’t great. Andrei wouldn’t use it anyways because it misses out on DRAM as he realized but he rightly points out it doesn’t even do the CPU well.

I would be interested in any evidence you have for this? How exactly do you know power metrics is badly modelled and inaccurate? Why would the wall minus idle on a device with a battery be better.

I’m also a little surprised to hear that the M2 is using more power than previously reported, and that the M4’s power usage is cause for concern.

The wall or ideally PMIC/VRMs is how the power is actually measured in all the tests everyone points to for years from Geekerwan and Andrei or Anand.


I’m honestly surprised people took it seriously, it’s nearly irrelevant. We basically have (some) Apple fans doing their own version of AMD/Intel guys reading HwInfo but it’s even worse because this is a game Apple and mobile vendors do better on.

I also want to point out that power metrics doesn’t even claim to measure DRAM anymore since a major MacOS update IME, which is funny. Even if it were accurate, it would be near worthless.

AMD and Intel’s modeling are better without a doubt but still not to be trusted at all and their fans still play games on this note.


Again, this is so funny because Apple fans (or just anyone opposed to AMD/Intel sloppiness like myself) should not want to play the game of excluding DRAM and power delivery or package losses!


So yes, it’s better because it reflects dependent variables to the chip itself (DRAM access rates) or just directly linked like again, DRAM and the power delivery setup.

It’s honestly profoundly annoying.

 

[Jimmyjames]

Yes, the wall or someone testing the actual VRMs/power delivery is the best way to do this. That’s what Anandtech did for years with measuring A series power dominating rivals, you measure everything on down from the SoC to the package and DRAM to VRM/power stuff. These are all part of the product in a meaningful way, like how power costly DRAM accesses will vary based on L2 and SLC cache amounts, stuff like that.

Apple's M1 Pro, M1 Max SoCs Investigated: New Performance and Efficiency Heights

www.anandtech.com

And well (see power section) powermetrics is modeled internally and it ain’t great. Andrei wouldn’t use it anyways because it misses out on DRAM as he realized but he rightly points out it doesn’t even do the CPU well.

The wall or ideally PMIC/VRMs is how the power is actually measured in all the tests everyone points to for years from Geekerwan and Andrei or Anand.


I’m honestly surprised people took it seriously, it’s nearly irrelevant. We basically have (some) Apple fans doing their own version of AMD/Intel guys reading HwInfo but it’s even worse because this is a game Apple and mobile vendors do better on.

I also want to point out that power metrics doesn’t even claim to measure DRAM anymore since a major MacOS update IME, which is funny. Even if it were accurate, it would be near worthless.

AMD and Intel’s modeling are better without a doubt but still not to be trusted at all and their fans still play games on this note.


Again, this is so funny because Apple fans (or just anyone opposed to AMD/Intel sloppiness like myself) should not want to play the game of excluding DRAM and power delivery or package losses!


So yes, it’s better because it reflects dependent variables to the chip itself (DRAM access rates) or just directly linked like again, DRAM and the power delivery setup.

It’s honestly profoundly annoying.

I’m enquiring how you know it’s inaccurate, not just to restate the assertion.

 

[Artemis]

It's actually pretty easy to get an accurate report of total system power usage when you are on battery. You can get things like instant amperage and voltage from ioreg from the macOS command line. I do this all the time from a command line tool that I wrote. For example:

Current charge 4391 mAh
Full charge capacity 4598 mAh
Charge at 95.400% capacity
Design capacity 4563 mAh
Current full charge at 100.700% design capacity
Cycle count is 31
Battery temperature 79.07 °F (26.15 °C)
Instant Amperage -325 mA
Battery voltage 13.01 V
Discharging with -4.228 Watts
Power adapter disconnected
Current capacity 100%
State of charge 95%
Time Remaining 13:30

To run in clamshell mode on battery with an external monitor, keyboard, and mouse you can use pmset:

pmset disablesleep 1

Well, that’s not going to be fully accurate for one thing, not for what we want. You need a proper sampling rate for many tests because power fluctuations happen so rapidly across the subtests and idling. This absolutely won’t capture that because the polling isn’t accurate, or fast enough, and you can’t reconstitute the waveform with any accuracy.


It’s a Grover-Nyquist theorem application but for power variations in a platform.

Fwiw this has been the understood model of power (either external from the wall or to the internal power delivery) along with the sample rate issue for years — Andrei and Anand have spoken about this, and Geekerwan the guy who is now still doing the best measurements around talks about this — see the closed captions.

 

[Artemis]

I’m enquiring how you know it’s inaccurate, not just to restate the assertion.

If you would read, you’d see I did explain not only that via a link but why it’s irrelevant even if it were accurate, because it doest measure enough. It’s not a slight to Apple but it is an irksome tech misunderstanding reminiscent of AMD/Intel fans doing similar stuff.

And RE: DRAM, in the review Andrei used PM when they claimed to measure it — they no longer do, I would know because I had a Mac and they removed that in PM, forget which update. Now it’s just CPU, GPU and ANE power.


Also: I linked to Andrei’s article. It’s all over the place. Package Power = PowerMetrics

It’s not useful other than at the limit and this is still true today and he maintains it fwiw.

 

[Artemis]

I would be interested in any evidence you have for this? How exactly do you know power metrics is badly modelled and inaccurate? Why would the wall minus idle on a device with a battery be better.

I’m also a little surprised to hear that the M2 is using more power than previously reported, and that the M4’s power usage is cause for concern.

I don’t think we know enough yet. I’d like to see a normal Spec or GB6 test. If it works in an iPad it can’t be actually bad. They are pushing it though.

 

[dada_dave]

Yes, the wall or someone testing the actual VRMs/power delivery is the best way to do this. That’s what Anandtech did for years with measuring A series power dominating rivals, you measure everything on down from the SoC to the package and DRAM to VRM/power stuff. These are all part of the product in a meaningful way, like how power costly DRAM accesses will vary based on L2 and SLC cache amounts, stuff like that.

Apple's M1 Pro, M1 Max SoCs Investigated: New Performance and Efficiency Heights

www.anandtech.com

And well (see power section) powermetrics is modeled internally and it ain’t great. Andrei wouldn’t use it anyways because it misses out on DRAM as he realized but he rightly points out it doesn’t even do the CPU well.

The wall or ideally PMIC/VRMs is how the power is actually measured in all the tests everyone points to for years from Geekerwan and Andrei or Anand.


I’m honestly surprised people took it seriously, it’s nearly irrelevant. We basically have (some) Apple fans doing their own version of AMD/Intel guys reading HwInfo but it’s even worse because this is a game Apple and mobile vendors do better on.

I also want to point out that power metrics doesn’t even claim to measure DRAM anymore since a major MacOS update IME, which is funny. Even if it were accurate, it would be near worthless.

AMD and Intel’s modeling are better without a doubt but still not to be trusted at all and their fans still play games on this note.


Again, this is so funny because Apple fans (or just anyone opposed to AMD/Intel sloppiness like myself) should not want to play the game of excluding DRAM and power delivery or package losses!


So yes, it’s better because it reflects dependent variables to the chip itself (DRAM access rates) or just directly linked like again, DRAM and the power delivery setup.

It’s honestly profoundly annoying.

I’m enquiring how you know it’s inaccurate, not just to restate the assertion.

Looking back at the article and from my memory of Andrei's twitter was that powermetrics was mostly okay but could occasionally give odd results. Like CB23 where ST results were higher in software than in wall power, 502 too. However, it is not entirely the case that powermetrics doesn't measure DRAM and package, it actually used to. It doesn't anymore. Also, Geekerwan actually uses a separate set of power measurement APIs that may underpin powermetrics. @leman uses the same ones. They no longer report DRAM or package power either. As far as I know, Andrei was one of the few to actually try to measure iPhone power directly, physically. As far as I know, no one has done that for an Apple computer except wall power which does measure more than just CPU power. That can be what you want, subtract idle power, and you get rid of most of it, but not everything.

Well, that’s not going to be fully accurate for one thing, not for what we want. You need a proper sampling rate for many tests because power fluctuations happen so rapidly across the subtests and idling. This absolutely won’t capture that because the polling isn’t accurate, or fast enough, and you can’t reconstitute the waveform with any accuracy.


It’s a Grover-Nyquist theorem application but for power variations in a platform.

Fwiw this has been the understood model of power (either external from the wall or to the internal power delivery) along with the sample rate issue for years — Andrei and Anand have spoken about this, and Geekerwan the guy who is now still doing the best measurements around talks about this — see the closed captions.

View attachment 29443

View attachment 29444

While I know you were responding to using the battery for power, you can up the polling rate for powermetrics, but I would agree that if you can get access to VRMs and the board and measure directly that is the best. Sadly no one spends the time to do that for Macs. May be harder to do for Macs? not sure, I've never done anything like that, probably just no one spends the time to do it since we're generally a small audience. Again I believe Geekerwan is simply using another software metric, @leman uses the same ones they do.

If you would read, you’d see I did explain not only that via a link but why it’s irrelevant even if it were accurate, because it doest measure enough. I don’t appreciate that response when I am in fact stating something many here didn’t know apparently, and it’s not a slight to Apple but it is an irksome tech misunderstanding reminiscent of AMD/Intel fans doing similar stuff.

And RE: DRAM, in the review Andrei used PM when they claimed to measure it — they no longer do, I would know because I had a Mac and they removed that in PM, forget which update. Now it’s just CPU, GPU and ANE power.


Also: I linked to Andrei’s article. It’s all over the place. Package Power = PowerMetrics

View attachment 29447

It’s not useful other than at the limit and this is still true today and he maintains it fwiw.

I would argue that most of those are sane with differences explainable by power losses between wall and CPU. Obviously the CB23/502 ST results are suspect. I remember Andrei being more impressed when he first started using though obviously it isn't perfect. As most software measurements have issues.

 

[Jimmyjames]

I don’t think we know enough yet. I’d like to see a normal Spec or GB6 test. If it works in an iPad it can’t be actually bad. They are pushing it though.

Are you sure you’re not mich7y?

 

[dada_dave]

Are you sure you’re not mich7y?

? I think he’s saying if the power envelope is fine for an iPad it should be okay. Geekerwan’s tests are under extreme cooling and power draws may be higher as a result. In other words he’s saying the M4 probably doesn’t draw as much power under normal conditions in SPEC/GB as Geekerwan measures.

 

[Artemis]

Looking back at the article and from my memory of Andrei's twitter was that powermetrics was mostly okay but could occasionally give odd results. Like CB23 where ST results were higher in software than in wall power, 502 too. However, it is not entirely the case that powermetrics doesn't measure DRAM and package, it actually used to. It doesn't anymore.

Yeah I explained this. But what you might have missed is, in the review Andrei was using the one that did, and it was still inaccurate:

“In multi-threaded scenarios, the package and wall power vary from 34-43W on package, and wall active power from 40 to 62W. 503.bwaves stands out as having a larger difference between wall power and reported package power – although Apple’s powermetrics showcases a “DRAM” power figure, I think this is just the memory controllers, and that the actual DRAM is not accounted for in the package power figure – the extra wattage that we’re measuring here, because it’s a massive DRAM workload, would be the memory of the M1 Max package.“

RE: power — the losses from the wall are probably around 10%. Doesn’t explain enough. I think the sampling rate is the second issue for tests like GB

Also, Geekerwan actually uses a separate set of power measurement APIs that may underpin powermetrics.

I will check back on this! I could be wrong. This has to be *very* recent though no? This was with the A17 that he did this? I think he was adamant about not doing that.

@leman uses the same ones. As far as I know, Andrei was one of the few to actually try to measure iPhone power directly, physically.

Well with the 15 it was still Lightning to USB-C monitoring, and that is what Geekerwan *was* doing too. Golden Reviewer is another one, his tests are imperfect but directionally seem accurate.

As far as I know, no one has done that for an Apple computer except wall power which does measure more than just CPU power. That can be what you want, subtract idle power, and you get rid of most of it, but not everything.

Yes, but you get rid of enough of it to count and the alternative is SO much dumber.

Like, you want the package/DRAM and power delivery. That’s a key variable that makes Apple silicon (or Qualcomm’s and I bet you MediaTeks later) etc so much more efficient in the “real world” and during active power besides lower idle draw. If you guys want to see what “CPU” and “package” power draw looks like when the others play that game, Apple will broadly still win (for 128B bus parts) but the others are going to end up much closer* because of the other elements.

It’s a bad hill to die on.


*I’ll post exactly what I mean in a minute.

While I know you were responding to using the battery, you can up the polling rate for powermetrics,

Not enough afaict!

but I would agree that if you can get access to VRMs and the board and measure directly that is the best. Sadly no one spends the time to do that for Macs.

I think the wall is close enough but fwiw vrms would be ideal yes. That’s what those Qualcomm graphs are.

May be harder to do? not sure. I've never done anything like that. Again I believe Geekerwan is simply using another software metric, @leman can confirm.

It’s only very recently that this is true, just checked! Originally he was doing external sampling and probably still does for Android.

 

[Jimmyjames]

? I think he’s saying if the power envelope is fine for an iPad it should be okay. Geekerwan’s tests are under extreme cooling and power draws may be higher as a result. In other words he’s saying the M4 probably doesn’t draw as much power under normal conditions in SPEC/GB as Geekerwan measures.

Perhaps it’s my ignorance. I did not get the same message from the link as I did from the message here. I don’t see any basis for the statement that wall power is superior to powermetrics. Again though it may well be me. It’s been a tough day!

 

[Jimmyjames]

The parallel discussions on here and anandtechs forums are weird!

 

[Artemis]

Perhaps it’s my ignorance. I did not get the same message from the link as I did from the message here.

I don’t see any basis for the statement that wall power is superior to powermetrics. Again though it may well be me. It’s been a tough day!

This is some serious blind tribalism. It got even worse since he wrote the review — because there’s not even an attempt at DRAM power now. Disheartening but possibly unsurprising to see some Apple fans are becoming just as bad as like AMD and Intel guys now about accuracy.

 

[Jimmyjames]

This is some serious blind tribalism. It got even worse since he wrote the review — because there’s not even an attempt at DRAM power now. Disheartening but possibly unsurprising to see some Apple fans are becoming just as bad as like AMD and Intel guys now about accuracy.

It’s blind tribalism to say I may be wrong?

 

[dada_dave]

Yeah I explained this.

Yeah I'm struggling to keep up with conversation. You guys are writing a lot faster than me.

But what you might have missed is, in the review Andrei was using the one that did, and it was still inaccurate:

“In multi-threaded scenarios, the package and wall power vary from 34-43W on package, and wall active power from 40 to 62W. 503.bwaves stands out as having a larger difference between wall power and reported package power – although Apple’s powermetrics showcases a “DRAM” power figure, I think this is just the memory controllers, and that the actual DRAM is not accounted for in the package power figure – the extra wattage that we’re measuring here, because it’s a massive DRAM workload, would be the memory of the M1 Max package.“

True. Though validating the core power measurements is a lot harder and I think that's what @Jimmyjames is getting at.

RE: power — the losses from the wall are probably around 10%. Doesn’t explain enough. I think the sampling rate is the second issue for tests like GB


I will check back on this! I could be wrong. This has to be *very* recent though no? This was with the A17 that he did this? I think he was adamant about not doing that.

Well with the 15 it was still Lightning to USB-C monitoring, and that is what Geekerwan *was* doing too. Golden Reviewer is another one, his tests are imperfect but directionally seem accurate.


Yes, but you get rid of enough of it to count and the alternative is SO much dumber.

Like, you want the package/DRAM and power delivery. That’s a key variable that makes Apple silicon (or Qualcomm’s and I bet you MediaTeks later) etc so much more efficient in the “real world” and during active power besides lower idle draw. If you guys want to see what “CPU” and “package” power draw looks like when the others play that game, Apple will broadly still win (for 128B bus parts) but the others are going to end up much closer* because of the other elements.

It’s a bad hill to die on.

I'm not trying to die on any particular hill here. I quite like living.

*I’ll post exactly what I mean in a minute.

Not enough afaict!

Really? I managed to sample damn fast, though I can't remember was the limit was. It's in units of ms, but I don't know if 1 is the limit or its less or more. I don't know what the private APIs @leman uses can sample at.

I think the wall is close enough but fwiw vrms would be ideal yes. That’s what those Qualcomm graphs are.

It’s only very recently that this is true, just checked! Originally he was doing external sampling and probably still does for Android.

My memory is hazy but I remember Andrei around the M2/A15 timeframe fighting with the Chinese benchmarking community over proper power testing because they were getting slightly divergent results. Either Golden Reviewer or Geekerwan or both. I believe you're right that at the time they were all doing physical measurements, maybe it was sampling rate that was different, I can't remember. Bottom line he was getting better results for the A15 and it was a little contentious as to why. I guess it's just an aside as it isn't really relevant today.

Perhaps it’s my ignorance. I did not get the same message from the link as I did from the message here. I don’t see any basis for the statement that wall power is superior to powermetrics. Again though it may well be me. It’s been a tough day!

No worries. Personally I like the combination - wall power and powermetrics combined together as Andrei did. If we had power rails that be best, but again, as far as I know no one does that. Indeed for most PC reviews where they do that, and it's more rare than not, they use test benches with easy access to the boards. Obviously for Macs that's generally not an option.

This is some serious blind tribalism. It got even worse since he wrote the review — because there’s not even an attempt at DRAM power now. Disheartening but possibly unsurprising to see some Apple fans are becoming just as bad as like AMD and Intel guys now about accuracy.

It’s blind tribalism to say I may be wrong?

I think you guys are being a little unfair to each other.

 

[Artemis]

The parallel discussions on here and anandtechs forums are weird!

The anandtech guys think M4 is a failure which I don’t agree with on account of the recent scores that show it actually does have a 10% IPC uplift. The overall Apple IPC slowdown though is true, not that they could sustain the same exact pace though. But this might be a start of a reversal. What people miss is:

The M/A series is still on top in ST efficiency and E core efficiency (at the same performances vs competition). Even if they have been largely riding off the foundation built already and tweaking it, using new nodes, (pending the recent change) it’s still on top.

AMD’s Phoenix on N4 still uses 2.5x the power to match an M2 ST. The idle is crap.

Or — in the 8 Gen 3, with an X4 on N4P — it still uses 20-30% more power to just barely pass the (N5!) A14’s SPEC score. This stuff comes down to architecture and cache — muh node isn’t everything.

 

[dada_dave]

The anandtech guys think M4 is a failure which I don’t agree with on account of the recent scores that show it actually does have a 10% IPC uplift. The overall Apple IPC slowdown though is true, not that they could sustain the same exact pace though. But this might be a start of a reversal. What people miss is:

The M/A series is still on top in ST efficiency and E core efficiency (at the same performances vs competition). Even if they have been largely riding off the foundation built already and tweaking it, using new nodes, (pending the recent change) it’s still on top.

I dunno whenever a rival chip design uses a wider decode or something, their marketing department start talking about brand new architectures and clean sheet designs!

AMD’s Phoenix on N4 still uses 2.5x the power to match an M2 ST. The idle is crap.

To be fair their idle isn't as bad on laptops, on desktops it's a disaster though I am not sure why ... often when I've seen idle power measure they have a discrete GPU plugged in (notebookcheck does pretty good wall power measurements though unfortunately CB R15, but they often use an Nvidia 3090/4090 in their test setup), sometimes those can draw large power under idle. That might be the cause. But the laptop idle was reasonable.

EDIT: Well for the 7940HS some of them were reasonable, some tested laptops not so much (according to Notebookcheck). Idle power:

Model	CPU	GPU	RAM	Value
Geekom A7	R9 7940HS60 W / 35 W	Radeon 780M 2800 MHz	32 GB	4.81
Minisforum Venus Series UM790 Pro	R9 7940HS65 W / 0 W	Radeon 780M	16 GB	6.05
Asus ROG Zephyrus G14 GA402XY	R9 7940HS80 W / 80 W	NVIDIA GeForce RTX 4090 Laptop GPU 125 W 1455 / 2250 MHz 16 GB	32 GB	6.16
Framework Laptop 16	R9 7940HS79 W / 54 W	Radeon RX 7700S 100 W ? / 912 MHz 8 GB	32 GB	13.9
Asus TUF Gaming A16 FA617XS	R9 7940HS114 W / 105 W	Radeon RX 7600S 95 W 2207 / 2000 MHz 8 GB	16 GB	17
Asus ROG Flow X13 GV302XV	R9 7940HS65 W / 60 W	NVIDIA GeForce RTX 4060 Laptop GPU 60 W 1520 MHz 6 GB	16 GB	18.8
Razer Blade 14 RTX 4070	R9 7940HS88 W / 80 W	NVIDIA GeForce RTX 4070 Laptop GPU 140 W 1980 / 2000 MHz 8 GB	16 GB	19.9
Asus TUF Gaming A17 FA707XI-NS94	R9 7940HS80 W / 80 W	NVIDIA GeForce RTX 4070 Laptop GPU 140 W 2030 / 2025 MHz 8 GB	16 GB	20.1

I'd say it was a discrete graphics problem except for the Asus ROG seems fine. Other AMD processor showed better or worse idle power, but it seems if the laptop maker cares they can get it down.

Or — in the 8 Gen 3, with an X4 on N4P — it still uses 20-30% more power to just barely pass the (N5!) A14’s SPEC score. This stuff comes down to architecture and cache — muh node isn’t everything.

Yeah ... actually my main concern is not with the single core performance figures but the multicore performance figures. Like I get how 12 P-cores clocked the way they do can draw 70-80W and I can see how the upclocked 12 cores will only be about 10% better for almost 2x power draw than the next tier down of processors. I finally worked all that out myself. That I get. But why the base multicore performance figures of the next tier down are so bad is a mystery - like 1100 in CB 24 and not great in GB6 either. With 12 P-cores clocked they way they are, they should be blowing M3 Pros and M2 Pros out of the water in terms of raw performance especially when all cores are turned on even if that burns more power so perf/W wouldn't be as great. I understand the Apple's E-cores are good, very good. But they ain't THAT good. So there's still a missing piece here that I don't understand and I'm not alone in these forums in confusion about Qualcomm's multicore performance characteristics. 12 M1/M2 like P-cores should be more performant when working in concert and they aren't and that's just bizarre.

EDIT2: oh shit, I'm tired you were writing X4, I was reading Oryon. Sorry. But the above still holds for Oryon. Very confusing.

 

[Artemis]

Yeah I'm struggling to keep up with conversation. You guys are writing a lot faster than me.

True. Though validating the core power measurements is a lot harder and I think that's what @Jimmyjames is getting at.


I'm not trying to die on any particular hill here. I quite like living.


Really? I managed to sample damn fast, though I can't remember was the limit was. It's in units of ms, but I don't know if 1 is the limit or its less or more. I don't know what the private APIs @leman uses can sample at.

My memory is hazy but I remember Andrei around the M2/A15 timeframe fighting with the Chinese benchmarking community over proper power testing because they were getting slightly divergent results. Either Golden Reviewer or Geekerwan or both. I believe you're right that at the time they were all doing physical measurements, maybe it was sampling rate that was different, I can't remember. Bottom line he was getting better results for the A15 and it was a little contentious as to why. I guess it's just an aside as it isn't really relevant today.

It was the GPU. Apple had a firmware difference where it would recharge and cycle from some battery % point, or like after 30 minutes of gaming was when you could get steady state. Golden Reviewer said his measurements were too low. I trust Andrei.

No worries. Personally I like the combination - wall power and powermetrics combined together as Andrei did. If we had power rails that be best, but again, as far as I know no one does that. Indeed for most PC reviews where they do that, and it's more rare than not, they use test benches with easy access to the boards. Obviously for Macs that's generally not an option.




I think you guys are being a little unfair to each other.

I think for most reviewers something like a constant test and the wall might work, also just doing battery drain tests and measuring performance is good. Those, and then idle ones for video.

But what I think is often most telling is automated web browsing tests.

 

[dada_dave]

It was the GPU. Apple had a firmware difference where it would recharge and cycle from some battery % point, or like after 30 minutes of gaming was when you could get steady state. Golden Reviewer said his measurements were too low. I trust Andrei.

Yeah now that I you bring it up I do remember the GPU as well, but I thought it was the CPU too, because I remember Andrei arguing with them whether or not the A15 used more or less power in CPU - because Avalanche was very similar to Firestorm Andrei contended that while it drew more peak power, the tweaks they made and the extra performance made up for it and he was saying its perf/W was still better, whereas Golden Reviewer was saying it was worse. But maybe I'm getting confused and it was just the GPU, that was several years ago now.

I think for most reviewers something like a constant test and the wall might work, also just doing battery drain tests and measuring performance is good. Those, and then idle ones for video.

But what I think is often most telling is automated web browsing tests.

How so for automated web browsing tests?

 

[Artemis]

It’s blind tribalism to say I may be wrong?

Fair enough, I am just kinda tired of hardware power shenanigans. You (might?) really have no idea how bad it still is. Tribalism is fine to a degree, but IMO it’s important to be accurate.

Yeah now that I you bring it up I do remember the GPU as well, but I thought it was the CPU too, because I remember Andrei arguing with them whether or not the A15 used more or less power in CPU - because Avalanche was very similar to Firestorm Andrei contended that while it drew more peak power, the tweaks they made and the extra performance made up for it and he was saying its perf/W was still better, whereas Golden Reviewer was saying it was worse.

Yes there was that too, he didn’t do a thorough enough job. The A15 used N5P which showed 5% more performance ISO-power, but Apple took a bit more than that, and of course that’s only for the die, the DRAM didn’t gain from that and then it’s running a bit faster.

What made the iPhone 13 more efficient yet still was four things, two related to ST efficiency, others just total platform for the mobo.

1: 50% more L2 for P cores. 8 -> 12MB. Shared L2 is just really awesome and a more efficient way to allocate SRAM, insofar as you can access more than half at a given moment (and you can with Apple’s, like 2/3 iirc) for ST, and make it kind of like a faster L3, but this only really works well if you have a big L1 I think to buffer latency.

The second thing on that same note and bigger was the SLC. SLC went from 16 to 32MB. The last two things just broadly are the X60 Modem and the upgraded E Cores which were a genuinely big boost above the A14.


The display was better too but that’s separate

 

[Cmaier]

And my M4 ipad pro is all set up! Took a long time to download everything off of icloud, and i had to jump through some hoops to get my work MDM set up right. But it all seems to be working.

Haven’t done much yet other than squeeze my apple pencil pro. Oooo. clicky.

Hmm. volume buttons are now reversed? And to make the tool pallet go away after clicking pencil you have to click it again, looks like. After choosing a tool, you then get more options in the pallette (line width, color. etc.)