I always thought that the hardest part must not be dealing with the complexity of average x86 instructions, but rather having to deal with the possibility of ever encountering one of those few instructions that are too complex, however rare those may be now.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
[CPU only] Apple M1/2(Max/Ultra) (TSMC 5nm) vs AMD Zen 4 (TSMC 5nm) - Technical Analysis
- Thread starter exoticSpice
- Start date
- Sort by reaction score
I suspect the hardest part is decoding multiple x86 instructions from the same thread in parallel. The special decoder which can emit a long stream of µops for the rare complex ones just needs an extra-large µcode ROM; it's not doing anything special other than that.
Hey mi7chy, nice quote you have, "MR is a graveyard of misinformation". It seems you are also part of this problem.
oh and mind you every internet forum as misinformation.
Any fixed-size VLA block can be decoded in O(1), you just need to throw enough die area and power at it. Definitely not trivial but doable.
Part? One of the core perpetrators more like.
And how did you determine that? For a truly "Apples to Apples" comparison you must have the same chassis, with exact same components on all systems, & ideally same software & OS to compare them with. Now granted the latter is unrealistic because Mac OS is one of the advantages Apple has, but then I guess you also saw benchmarks of zen4 mobile chips being used with LPDDR5 already
I don't trust software reading for power consumption, they're never really that accurate especially at lower loads so anyone claiming Apple's massive lead purely based on software needs better evidence!
Thats why do a battery draining test with the same apps open. Best way to judge for efficiency and take into account the battery size and screen size/res/brightness.
Simple and easy.
Sure, but there's no tests for zen4 (mobile) let alone a laptop using LPDDR5 w/zen4 cores atm. If we're talking about previous gen Intel/AMD chips then yes I partially agree with the efficiency claim, though again they were disadvantaged by an inferior node.
Wouldn't it technically be a 'seedbed' of misinformation? Maybe it's only a graveyard for misinformation from the perspective of the beholder
Same chassis? Nonsense. All you need is equivalent loading and the ability to measure, either directly or indirectly, how much energy was consumed to perform a task. If both chassis keep the processor cool enough to avoid thermal throttling, you're making a valid comparison.And how did you determine that? For a truly "Apples to Apples" comparison you must have the same chassis, with exact same components on all systems, & ideally same software & OS to compare them with. Now granted the latter is unrealistic because Mac OS is one of the advantages Apple has, but then I guess you also saw benchmarks of zen4 mobile chips being used with LPDDR5 already
I don't trust software reading for power consumption, they're never really that accurate especially at lower loads so anyone claiming Apple's massive lead purely based on software needs better evidence!
Same software & OS? Also nonsense. How can we even run the exact same OS on a M1 system and a PC, and why would that be a useful comparison? Almost nobody buys a M1/2 Mac to run anything other than macOS. (Technically you can use Linux, but the Asahi Linux project is far from complete in its support of M1 power management and other SoC features.)
You don't trust software readings of power consumption and somehow you magically know that they're inaccurate at low load? Okay, why should I care what you think? Why do you suppose Intel, AMD, and Apple put power and temperature sensors into their designs for software to read out? Hint: it isn't for marketing, and gross inaccuracy would defeat the purpose.
And who's basing this purely on reading on-die sensors? People have done bigger-picture tests. That's what went on in the video review someone linked a page or two back. Same python script running the same battery rundown test on different machines / operating systems, and the M1/M2 Macbook batteries lasted far longer than the Intel/AMD laptops. It wasn't even close, even where battery capacities were similar.
Apple is way ahead on power efficiency, and it's by ratios which make nitpicking about chassis and so forth look like desperation from fanboys who don't want to admit reality. This is true whether you're looking at micro level results (SoC power sensor data) or macro/system level (how long does the machine take to run down its battery).
And how are you measuring energy consumption genius on different laptops with different screens, memory SSD, wifi & even 5G sometimes? When you don't have access to measure SoC power consumption directly the next best thing is to use the same chassis kinda like AT did with MS' surface pro! Without eliminating the obvious variables you're not counting the chip's efficiency you're just comparing which "laptop" is more efficient!
The only ones desperate are Apple cult members, probably like you, who'd disparage better testing methodology to delve deeper into the efficiency of Mx but hey keep that magic pixie dust flowing 🥱
Of course they are less reliable than hardware measurements! Besides have you opened hwinfo on an AMD machine recently & checked the snapshot CPU polling interval option or changed the polling interval from 2s to say a second or lower? When the CPU is entering low power state & the software utility is measuring its power consumption the CPU can be forced to enter higher power states, this depends on the software as well as OS. So you're not really getting the best picture when it comes to energy efficiency. The same goes for peak power consumption - no software is 100% accurate in that regard. This is why margins at smaller loads of 10W & under can skew the results massively in one direction or another depending on the platform you're testing it on & your testing methodology.
What bigger picture tests & what do they achieve?
Yeah keep telling yourself that with zero evidence to backup your claims.
The only ones desperate are Apple cult members, probably like you, who'd disparage better testing methodology to delve deeper into the efficiency of Mx but hey keep that magic pixie dust flowing 🥱
How come the node issue only comes when comparing Apple vs AMD/Intel but not when comparing Intel vs AMD.
When it comes to x86 its Intel's fault for using a older node and AMD is cheered for using a better node. Anyway, Apple M chips and AMD/Intel chips won't be on the same node till 2024 that is N3.
Many battery tests show that M2 is more efficient than 6800U. This test normalized battery size, app and brightness:
source:
Again at the risk of repeating myself - this isn't how you compare the chip's efficiency.
It was for a long time when AMD lagged Intel by 2-2.5 nodes around the Bulldozer era.
Last edited:
ok then why is M2 MBP getting 4 hrs more battery life. If it was less efficient then it would get less than AMD and Intel.
Keep in mind the battery size, brightness, app used and the video played was the same.
Then it stated as fact that a mobile/laptop chip that the best node has the best efficiency. Of course design of the SoC/CPU plays a huge part too.
That is why the A16 is around 40% faster and also more efficient than the Snapdragon Gen 1 Plus even they are both based on the 4nm TSMC node.
Always an important question isn’t it? When Zen 2 made a meaningful performance/watt and process jump over Intel’s current options you heard very little about the process node being the main reason for it. A lot of celebration occurred regarding its performance and efficiency though.
M1 shows up and process is apparently the most important thing.
When the results are so heavily in favor of Apple, and when the comparisons are between broadly similar laptops, what's the point in splitting hairs? Rational people look at ratios of 2x (or more) and concede that Apple's ahead without insisting on arbitrary and impossible levels of rigor.
Sigh. What do you suppose is going on when you make a "software" power measurement? It's literally hardware: there are on-die current and voltage sensors, and software queries them.
If you're designing a modern SoC, you're going to put in a power management unit of some kind, and it needs data. The better the data, the better it does its job. That's why we end up with lots of on-die I, V, and T sense circuits. As a side effect it becomes possible to write tools which query the PMU for summary data.
If the AMD machine loses lots of power efficiency just because a sensor poll happens faster than once every 2s, that's a sign AMD has a power efficiency problem on light loads. Modern computers have all kinds of random things waking up and doing minor tasks at regular intervals. Apple Silicon appears to be best-in-class at minimizing power used for these kinds of things, and that is a legitimate advantage which you don't get to handwave away.
You're delusional if you think it hasn't been proven over and over that M1 and M2 are way more power efficient under both light and heavy loads.
Go watch that youtube video @Xiao_Xi posted on Monday. It's quite thorough and its creators put some thought into methodology. I don't agree with all their explanations, but that's OK, I think they did some good work and the results speak for themselves.
Or read this:
There's at least one SPEC multithreaded test (503, aka 503.bwaves_r) where AT measured at-the-wall perf/W of M1 Max at over 6x that of the i9-11980HK. No amount of special pleading about backlights or chassis or whatever other excuse you come up with is going to argue away a ratio of 6x. The other data from that AT test isn't quite so extreme, but 2x to 4x is common. There's still no way you can handwave that away. And while the 11980HK is now out of date, I doubt Intel's successor chips (or AMD's competition) are even as much as 2x as efficient as the 11980HK.
Face it, if there's anyone who's being a cult member here it's you. Your posts on this reek of cognitive dissonance. Your deeply held belief that Apple can't possibly be dramatically better at anything is being severely threatened by reality and you aren't handling that disconnect very well.
Oh my god. Do you think insisting on more direct testing methodology with a 100% controlled environment makes you somehow sound more science-y, or rigorous? Don't ever read a medical paper then, you'll find that all of them have zero evidence to backup their claims if that's where you're setting the bar. In the real world you can't expect everything to be the exact same. And yet you can make valid deductions with the data you have.
Or a particle physics paper, hah. Literally zero direct measurements in that field, and somehow the science keeps advancing.
The differences in efficiency being discussed here are huge. Even with the biggest error bars you could possibly put around your data, they wouldn't overlap.
Don’t run a LC0 GPU benchmark or a Stockfish CPU benchmark, because Apple M1/M2 Ultra will be killed.
Locuza has written about the evolution of Zen. It may help to better understand the benchmarks and reviews about Zen 4 that should be published today.
locuza.substack.com
Zen evolution: A small overview
I feel like the rabbit in Alice in Wonderland who is always short on time…
locuza.substack.com
If 7950x leaked benchmarks turn out to be true even at 65W it destroys M1 Ultra on both single and multi-core. Strong productivity performance but not as fast as fat cache 5800x3D for gaming.
https://itvision.altervista.org/zen4-ryzen-7950x-leaked-review.html
https://itvision.altervista.org/zen4-ryzen-7950x-leaked-review.html
Could be because of Windows.
Try the same test with Linux, which is faster than Windows as an OS.
Linux generally has worse battery life. Sleep and idle states are also finicky on Linux too
Yep the 7950x is a beast. it does beat the M1 Ultra thanks to it's insane clocks.
I would use Geekbench for Single Core tho as AMD does the same.
Cinebench is not 100% optimised for ARM even though it's an Apple Sillicon binary.