M3 Max power consumption - it's remarkably efficient

[danwells]

I've been looking for some numbers on power consumption under load since the M3 line came out. I've finally found a few useful ones, and they're a LOT better than any comparably fast Windows laptop (expected, but confirmation is nice)...

A couple of YouTubers have finally benchmarked some photo and video applications (Lightroom, Photoshop, Resolve), and the speed is excellent (as we've known - ~150% of M2 Max on most tests, and quite a bit better than any laptop PC, even plugged in - the Mac will run full speed on battery, and NO PC will) - but they also looked at battery life and power consumption. A lot of this is MaxTech, and their commentary is as hyperbolic as usual, but it's their numbers (probably not faked) that are interesting.

They showed both the CPU and GPU able to crank up to 50 watts or a little more (each), and they had 42% battery remaining after 3 hours of benchmarking. If we take benchmarking as about as tough as "normal" creative workloads get (I say "normal" because somebody's inevitably going to try exporting a feature film on battery), we can count on at least 4-5 hours on battery in most interactive editing workflows, often more. Nice to know that number, which is one of the really important ones for users of this kind of laptop. We know it gets 22 hours with the CPU mostly shut down, watching movies on the decoder chip, but who buys this thing for that? We know it gets 15 hours loafing along in Word and Safari, which is useful to know because most people who use Lightroom, Capture One, Resolve, etc. ALSO spend significant time in Word and Safari - but that doesn't answer "how does it do when pushed"?

We finally have an answer to that - 4-5 hours in tough editing workflows, probably better in more contemplative stuff which stresses the processor, but only intermittently. That's a great number. and there is no PC that comes close (most gaming PCs and laptop workstations have a hard time managing 4-5 hours in WORD). All day (~8 hours) in photo or video editing is possible, depending on what you're doing.

You can't come close to the CPU performance with anything else that fits in a backpack - the fastest PC laptop chips are 30% slower, while using more than twice the power. GPU tests are closer, and some high-power laptop GPUs can beat it even in fair, actually cross-platform tests (they can beat it by more in gaming tests, because Apple's GPUs have always been optimized for creative use instead of gaming, and even Mac-ported games tend to use Rosetta, OpenCL or both). A RTX 4080 laptop will just barely beat the M3 Max, and a 4090 will beat it by ~25%, both while drawing 175 watts instead of 50.

NotebookCheck managed to push it harder with power virus workloads (it looks like they used Witcher 3 pushed really hard, although I'd have liked more clarity - games can drain batteries in ways that very few other apps can). They managed to drain the battery in 1 hour, 11 minutes, which is comparable to the best gaming laptops - but the gaming laptops throttle on battery, which the Mac doesn't. They got the M3 Max up to 140 watts in a burst, while gaming laptops (of comparable performance - 20-30% less CPU, but 20-30% more GPU) run as high as 270-280 watts in VERY short bursts before throttling to avoid catching fire.

The other implication of this is what performance vs. power will look like on the M3 Ultra, and if there's an M3 Extreme. The Ultra should be running as fast as kilowatt-plus PC workstations and gaming rigs, while consuming ~200 watts under load, bursting to ~300. If Apple makes a M3 Extreme (quad Max), CPU performance will be better than ANY PC workstation (with possible exceptions for some extremely parallel, generally multi-user workloads - the Extreme will beat the world in DaVinci Resolve, but maybe not in SQL server where 100 cores are useful), with GPU performance only approached by Titans and Multi-GPU rigs (while running at 400-600 wats instead of as much as a couple of kilowatts).

This thing is crazy fast, but also crazy efficient...

 

[ArkSingularity]

That kind of battery life is just incredible to me. I was very curious how the Max would do compared to the Pro, seems like both are doing quite well!

 

[TheRealAlex]

I would love to see 12 Core M3 Pro Vs. a 14 Core M3 Max Vs a M2 Pro. Doing Some Video editing, Photo Editing. And Im guessing the M3 Pro would have the BEST battery life of all. Thanks to a Smaller Die than the M2. And less power hungry than the M3 Max. So it's the perfect power to performance sweet spot.

 

[tekksan]

This guy did a test with the M3 Pro vs M3 Max.

He didn’t interpret the battery results correctly though. A commenter corrects him.

TL;DR the M3 Pro got about 25% better battery life in his battery test.

 

[danwells]

Interesting stuff - M3 Max being consistently ~150% of the M2 Max (which it is outside of a couple of tests that are really just running the decoder) is pretty incredible for one generation.

What I bet we're seeing in the M2 Ultra/M3 Max comparison is "how parallel is the task"? If you have no (or very limited) scheduler losses, the M2 Ultra will be almost exactly twice as powerful as the M2 Max, since it's literally two M2 Maxes glued together with some supercomputer epoxy (2.5 TB/s is an insanely fast interposer, of the type normally only found on a supercomputer - ordinary dual-processor computers use links of 50-200 GB/s). With average scheduler losses, the M2 Ultra seems to be about 1.5x as fast as the M2 Max or a little better. At the opposite extreme, the M2 Ultra is exactly as fast as an M2 Max on a single-core workload.

On an extremely parallel task that runs on P-cores, the M3 Max (assuming the 12 P-core version) should be about 1.725x as fast as an M2 Max (1.5x the number of P-cores, 1.15x the per-core speed). Any scheduler losses detract from that theoretical speed, until a single-core task is only 1.15x as fast as an M2 Max. Average 1.5x seems pretty reasonable in workloads that are neither fully parallel nor single-core.

The other one that caused me to gulp loudly was the Blender test that even came close to a RTX 3090 DESKTOP... Yes, we've been seeing speeds close to big MOBILE cards in some benchmarks, but anything other than an Ultra coming close to a 300 watt desktop card is new, even if the desktop card IS a generation old, and the test is one that loves Apple Silicon.

 

[danwells]

He didn't mention RAM, which matters for some tests (or which core count of the M3 Max he had - his results suggest the 16 core), and he was using a 14" machine. His fan noise test is invalid for the 16" version (which is reputed to be MUCH quieter).

Personally, I wouldn't put a M3 Max in the 14" chassis. It may be a remarkably efficient 100 watt chip that is trouncing gaming laptops and mobile workstations that draw twice as much power, but it still peaks at or above 100 watts, and its cooling needs must be respected. It's so fast and so efficient that it'll run under 20 watts a lot of the time even in Photoshop and friends, but it needs to be able to peak much higher, and the 14" model can get constrained on peaks.

A typical creative workflow (other than long batch exports and renders) involves a lot of time slightly above idle, punctuated by times at full power when it's running a filter, a merge or something else. A fast machine will be a lot more responsive, but it needs to be able to handle those peaks, and I'm much more confident in the 16" than the 14"...

 

[Beau10]

My 16" unbinned max has been great. Workflows my old 14" base M1 Pro could last only about 5 hours on (83% battery life, so about 6 new), it will go about 12 - easy all day use. I only charge it to 85% w/ AlDente as that is just about sufficient to have it survive most of the day without dipping too much into the sub 20% red area.

That said, you can still kill it quick. I played a session of BG3 at max settings a couple nights ago and it went from 75% to 5% in about 45 minutes. Total lap heater. Didn't have a power meter on but that means it was pumping out something like a consistent 70 watts.

 

[Kingcoherent]

These are very valuable benchmarks, thank you for pulling them together!

 

[Onimusha370]

Nice numbers! I was actually really surprised to see how much power the non-binned M3 Max could draw when needed. I saw 132 watts at its peak while doing some de-noising in LRCC. I'd imagine thats roughly 120 watts from the CPU/GPU/neural engine combined (once you remove power consumption of a full brightness screen). That's higher than my old rule of thumb for the M1 Max, which was 30W CPU and 60W GPU at full tilt. I guess we're now playing with 4 extra performance cores (roughly 45W CPU?) and 8 more GPU cores (roughly 75W GPU?).

I might download one of the power usage monitors and try and get more accurate results, but its no wonder the fans seem to spin up a bit more on the M3 Max series than the M1 Max - it seems the M3 Max has really moved up into another class (of both performance and power usage) compared to the M1 series when really pushed. Good trade off imo given how much headroom apple left themselves with on the thermal side of the M1 chips.

 

[danwells]

You can kill anything quick in a game. They are semi-unique workloads that put a sustained strain on the CPU and especially the GPU. No other type of interactive software comes close (batch renders and the like can, and VR experiences almost certainly will once they become more common - but VR is, at least mostly, a type of game).

There is very little in, say, photo, video or music editing that sustains that type of workload, other than batch jobs like importing, rendering or exporting. When you're editing interactively, you spend a lot of time deciding what to do next, then you apply an action and the processor spins up to do that action, then you look over the result and choose the next one. The reason for a very fast computer is so those intermittent actions are smooth (move a slider or change an edit point, and you can see the results), and to cut the time for big batch jobs. You aren't interacting directly with the processor at full power on the same continuous basis you are in a game.

 

[TheRealAlex]

All I care about is my M3 Pro 12 Core, plays DOTA 2 and StarCraft 2, and World War Z at Ultra with 100+ FPS Unplugged.

 

[Appletoni]

You can kill anything quick in a game. They are semi-unique workloads that put a sustained strain on the CPU and especially the GPU. No other type of interactive software comes close (batch renders and the like can, and VR experiences almost certainly will once they become more common - but VR is, at least mostly, a type of game).

There is very little in, say, photo, video or music editing that sustains that type of workload, other than batch jobs like importing, rendering or exporting. When you're editing interactively, you spend a lot of time deciding what to do next, then you apply an action and the processor spins up to do that action, then you look over the result and choose the next one. The reason for a very fast computer is so those intermittent actions are smooth (move a slider or change an edit point, and you can see the results), and to cut the time for big batch jobs. You aren't interacting directly with the processor at full power on the same continuous basis you are in a game.

Even chess can squeeze the M3 MAX out.
For the Stockfish chess engine we use all CPU cores.
For the LC0 chess engine we use all GPU cores.
For a more professional work with Stockfish, we would need at least +8 performance cores = 20 + 4 = 24 cores total.
MacOS, Windows 11 ARM and Linux are often used at the same time. With a lot software running on all at the same time. We are really happy about the 128 GB RAM which are obviously needed too. It's easily for a chess engine to take even TBs of RAM.
Metal 3 is also very important, hopefully we will see a much stronger Metal 4 and a lot of more neural cores.
8 TB SSD is the minimum for Software + Syzygy Endgame Tablebases. Hopefully they will sell a 16 TB SSD or 2x 8 TB SSD soon.

 

[djcristi]

all tests that reports powermetrics and not AC power measuring are wrong. powermetrics ar least on m3 max 14c does report a lower value than the real Soc consumption, i have explained it here

 

[danwells]

Would the best way to measure be a power meter between the wall and the laptop? Would that be accurate if the battery were fully charged (or is there still some buffering effect from the battery, where wall draw may not equal system power because some power is coming into or out of the battery)?

 

[danwells]

What do you run Stockfish on that has terabytes of RAM? A supercomputer???

 

[Appletoni]

What do you run Stockfish on that has terabytes of RAM? A supercomputer???

Sometimes.
But most of the time on a MacBook Pro 16-inch M3 MAX maxed out.

 

[danwells]

So 128Gb is good for what sounds like a REAL RAM-intensive process?

 

[Appletoni]

So 128Gb is good for what sounds like a REAL RAM-intensive process?

Yes but not only.
It's also great for other stuff.
For example Parallels Desktop with Windows 11 ARM can use up to half of that 128 GB RAM.
For example:
-I can use some GB RAM for macOS and run Stockfish chess engine analysis on macOS and in the same time I can use some GB RAM for Windows OS and play some top games there.
-In the same time I can also always let Safari with all tabs open, because I do a lot of working and learning stuff 24 hours a day.
-Most of the time macOS don't need to swap data from RAM to SSD and back, which is very good for many reasons.
-I can run 2 very different engines Stockfish and LC0 in the same time with 64 GB RAM per engine.
-I can use 128 GB RAM for my work.
-I also can run Linux and do stuff there.
-CrossOver works also great.

128 GB RAM are important because:
-I can run one high-intensive task and everything is fine.
-I can run multiple OS and multiple tasks in the same time and everything is fine.
-I can use it for a longer time (Years) compared to people with less RAM, because everything what we are doing on those devices is always increasing in size.
-I can due everything much faster.
-I can sleep quiet at night. 🥰

128 GB RAM sounds like a lot but in reality it isn't.
I hope to buy a 256 GB RAM MacBook from Apple as soon as possible.

 

[danwells]

256 GB is an interesting milestone, because you can't get there in any easy way right now (you need a very expensive Threadripper or Xeon). Macs use a bunch of memory channels, especially at the higher end (Max chips are 6 or 8 channel, while Ultras are 16 channel), but PCs are all dual-channel until you get to Threadrippers and Xeons.

Short of using RDIMMs, which conventional desktop CPUs don't support, you're limited to 32 or maybe 48 GB per DIMM, two DIMMs per channel and two channels. 128 GB is (relatively) cheap and easy, 192 GB is possible using 48 GB DIMMs (very latest CPU and motherboard both required) if you work around some compatibility issues, and >192 GB requires exotic hardware that supports >2 RAM channels and/or RDIMM.

In contrast, a 256 GB MacBook Pro is actually trivial to make, since Max chips already have six or eight memory channels. I don't want to know what Apple would charge for one, but it's as easy as replacing the 8x16GB RAM packages on the existing 16/40 M3 Max MBP with 8x32 GB. I'm not sure if a 512 GB MBP is possible, but it might be - it would require two RAM packages per channel, and I don't know if Apple Silicon supports that (a lot of architectures do). I also don't know if the extra RAM packages would physically fit (both in the MBP case and close enough to the SOC).

The high-end desktop (HEDT) chips that offer quad-channel RAM have moved WAY upscale. Intel used to offer chips that used a different socket from the standard desktop line, but without the Xeon name (they were more or less a hybrid between a traditional desktop chip and a Xeon-W), and Threadrippers used to start as low as $500 or so. It used to be that $500 for a processor and $250 for a motherboard got you four memory channels (often with eight DIMM slots) if you wanted them (the HEDT chips were pretty comparable in price to the top, super-overclockable conventional desktop chips), and a fair number of desktop PCs were built that way.

Intel no longer offers any quad-channel CPU/motherboard outside of full-blown Xeons, and a current-generation Threadripper STARTS at $1499 for the chip (with the versions with more cores ranging from $2499 on up to $10000 - the $1499 model is quite similar to a Ryzen 9 in performance, the real gain is expandability) and $699 for the motherboard. That's extremely niche - by the time you build a system around it, it would be VERY hard to come in under $5000, and a nice configuration will easily exceed $10,000 (a Xeon would be similar). There simply are no more $3000 quad-channel PCs, which once existed (at the time, 32 GB DIMMs were exotic if they existed at all, so there have never been reasonably priced ways of getting to 256 GB)...

I am not aware of any quad-channel LAPTOP that ever existed (being pedantic, other than the M1 Pro and M2 Pro). I'm reasonably confident nobody ever stuck a Threadripper in an oversize gaming laptop case and attached a motorcycle battery, at least not for commercial sale, but was there ever a weird laptop that used some Intel HEDT chip? Some of them were only around 100-125 watts, which is REALLY stretching what one might get into a laptop, but the Clevo (etc.) desktop-chip laptops sometimes use 95W chips... I seem to recall a 12 lb "laptop"/space heater that used DUAL 330W power adapters - could some version of it have had an HEDT socket?.

 

[macduke]

In the same time I can also always let Safari with all tabs open, because I do a lot of working and learning stuff 24 hours a day.

I've never felt more accurately described by another person, lol.

I was talking to my old boss about this a couple years ago because she thought she was a tab hoarder. I counted how many tabs I had open at the end of an especially busy week of working and learning stuff and it was nearly 500. She had like 70. I do web design, development and testing for my job as well, so that was across Safari, Chrome, Firefox, and Edge. I have no idea how people just use the same tab or three for everything. I'm always going back and forth between things, cross referencing, researching, comparing, testing, inspecting using dev tools, etc.

Safari has the most tabs by far because that's my default since it runs more efficient. Especially since a few years ago when they pretty much brought it up to feature parity with other browsers. I can keep way more tabs open in that than any other browser. Chrome has always been bad about using memory. But even so, typically my browser would go haywire and crash or become unresponsive before I would run out of the 64GB of memory on that old iMac. What's nice is so far the M3 Max has been using about 8GB less memory than my old iMac did. IDK why but I love it. I still think 64GB is more than I need, I only had that much in my old iMac because I upgraded it manually, which was possible, and it was cheap to do it. But for my next machine, around 2030, my house should be paid off and I'm gonna go ham with RAM lol.

 

[danwells]

Even with the unified memory (so the GPU's in there too), Apple Silicon has a reputation for running about 10% less RAM than Intel. I don't know if this is because code's more efficient, or what? I wouldn't surprised if, given its phone origins, it has very good memory management?

 

[djcristi]

powermetrics does not show memory power consumption, which is a lot, see my findings here

 

[xmach]

I was talking to my old boss about this a couple years ago because she thought she was a tab hoarder. I counted how many tabs I had open at the end of an especially busy week of working and learning stuff and it was nearly 500.

@djcristi How do you count the number of tabs open in Safari?
(on a related note, is there a way to do this also in the DuckDuckGo browser?