Transcription
Here I have the latest family of M4 MacBook Pros. M4 Max is the 16-in one, and then I have the M4 and the M4 Pro, but I forgot which one is which, so we're going to have to find out through the testing. I thought it would be a fun experiment. I also included my machine that I've been daily driving a lot for the past 2 years, my M2 Max MacBook Pro, just for comparison.
This is what a brand new machine looks like, nice and clean, and this is what a 2-year-old machine looks like. Look at all those keys; they're all shaved down. Apple really should do something about that coating. It's not even oil on my fingers; it just looks like it. Besides that, I treat this thing pretty nicely, so it's in good shape.
Now you've probably already seen some synthetic benchmark results, like Geekbench, for example, at which the entire M4 family just crushes. Now I'll be doing some real-world performance testing on these—some code compilation, running interpreted code, and some initial machine learning tests—so we can get an idea of what these chips are capable of. But I've also got a machine learning or AI deep dive video coming up, so don't miss that.
Oh, and since I'm just dipping my toes into these machines, I figured this would be a good starting point for comparisons. But give it a week or two, where I've had enough time with these machines to be able to really determine how they are in day-to-day tasks, especially in my own software development workflows.
Now before I get into the tests, just a couple of housekeeping things. Since I didn't do a dedicated unboxing video for the MacBook Pros, like I did for the Mac Minis, I usually set these up using Migration Assistant, which is a really good way to get things up quickly because I need to install a lot of software. As you know, these new machines use Thunderbolt 5 while the Pro and the Max do, but the M4 doesn't; that's still Thunderbolt 4.
I noticed a significant jump in transfer speed, and I just had to run iPerf right after that, sending packets from one machine to another. You set up a Thunderbolt Bridge in System Settings, Network, and you'll have different networks here like Wi-Fi. Thunderbolt Bridge is the one you would set up here. But if you do this, make sure you configure your IP address manually. Under Hardware, since these support really fast transfer rates, you want to make sure that you move this from standard to jumbo packets. When I did that, I got much faster speeds than Thunderbolt 4 could provide. We got up to 66 Gbps per second, which is crazy to be able to transfer that fast between the machines.
I've also set the M4 Pro and the M4 Max machines to high power mode when it's on battery, like right now, because I wanted to have the full performance benefits that that could give you. That's how I've been using my M2 Max for the past 2 years, with high performance mode. Does it have an effect on battery? Sure. But did I really notice it? No.
I'll be doing a separate video for battery comparisons and trying my automated overnight test on that to see how that affects it. One thing I do want to mention is if you're getting the M4 model, not the Pro or the Max, then you're not going to have the option for high performance mode; you're only going to be able to set it to low performance mode when the battery is significantly low. So keep that in mind.
One other thing we've seen through generations of these machines is the disk speed improved. So here we go; we've got pretty good scores around 6,200—right, 5,000 read. Now the M4 Max, wo, 7200 and almost 6,000. So better on this one. Interesting. I wonder if this machine is the M4 and it got lower SSD.
And here is the last machine. Oh, this one is lower. Okay, this one's only going to 3,300, right, and 2,600 read. They all have very different numbers, but also they have different storage capacities. I went all out with the M4 Max because that's going to be my machine for the next couple years, and that one has a 4 TB drive on it. This machine has a terabyte drive in it, and this machine has 512.
So the 512 GB machine is the slowest one, then the terabyte machine, and finally the 4 TB machine. I guess the bigger the drive, the faster they go. That's an interesting thing to know, considering Apple charges you quite a lot for these upgrades. Maybe it's not just space that they're charging you for; maybe it's the different NAND chips that are on the SSDs that they're charging you extra for because they're faster.
Here's my M2 Max from 2 years ago. This is what I've been using. It's about 3200, right, 5,000 read. That's a 2 TB drive, by the way. So what the heck is this?
Well, typically, I don't recommend working on a laptop outside in a bright day. First of all, the sun is going to be down on it, heating it up. I made videos on how heat affects the performance of the machines before. But if you are outside and it's a cold day, let's say, you're still going to be contending with the sun. These new MacBooks have a higher brightness display at 600 nits, but they're supposed to boost up to 1,000 if they detect a really bright environment.
The three new M4 MacBooks all have very close to the same kind of measurements. Right there, we've got a maximum of 567; that's on the Max. Those can be just slight variations in the screen size and the test. They're all within the same kind of range. This is pretty close to the 600 that's advertised; it's certainly higher than the 500, but the M2 Max is quite a bit lower than the 500 it's supposed to have—it's a 397. So significantly brighter displays on these new machines.
Now let's simulate some light hitting the sensors and see if we get a difference there. I don't know if it's just not a powerful light, but we didn't get much difference. All right, let's try the brightest light I have. Definitely can't see the screen here. Okay, 75%, I can see the screen pretty well, and 100%—whoa! There we go. Oh wow, look at that! 935 nits.
How did I do it? There is a setting here that says "Automatically adjust brightness." If that's not on, it's not going to do it; it's not even going to let you drag the slider to get that extra bump of brightness unless this is set to automatically adjust.
Now we're going to start building some code; it's about time, right? Since we're dealing with all Macs here, the best ecosystem to write macOS apps is Xcode. We're going to build two projects. One project has a lot of dependencies; it's called Xcode Benchmark, and I've ran it on the channel before. This is a test that includes over 70 popular CocoaPods as dependencies, so it's a pretty decent-sized build. There's C++, there's Swift, there's C, there's Objective-C, markdown, JSON—it's a pretty large project. Most projects are going to be smaller than this one, so this one takes a little bit of time to build. And of course, I've got my little friend here to help us out.
Even though this test does give us the time at the end, it's still more fun to use this one. I want to say its name, but I've had my videos flagged for saying its name. So if you know the name, put it down in the comments below.
We're ready to go here; when I push that button, these are all going to get kicked off at the same time, and let's go! Here we go; let the building commence.
Now we're also trying to figure out which one of these is the M4 and which one is the M4 Pro. The M4 Pro should be faster because it's got more cores. This is a multi-core build, so this might be a dead giveaway. If you have any guesses so far, you can put it down below.
Now the temperatures are holding steady—58° on that one, 47° on the big guy. Of course, the 16-in is going to spread heat a lot better than the 14-in machines. 58 on the other 14-in, 61 now. 884 over here on the M2 Max—whoa, M2 Max jumped up to 91 while the other ones are still staying kind of cool, especially that 16-in M4 Max. I’m starting to hear the fans kick in.
Which machine is it? The M4 Max is making noise, and it's done! Wow, it's the first one to finish at 81 seconds. Oh wow, let's take a look at the temperature while we still have it nice and toasty here. We're up to 38° during this test.
The machine to the left—oh, it's at 45°. See, that's pretty good for that 16-in. Slightly different profile here on this last machine—uh-oh! Does this one only have one fan? 43° on this one, 45°.
So notice this 14-in has two hot spots and this 14-in all the way on the right has one hot spot. I think we know which machine is which and how many fans it has. Both of those machines, the 14-in, got way hotter than the 16-in, but the 16-in kicked off its fans a lot sooner, and we have numbers, folks!
Now the M2 Max machine got 27 seconds in this test. You got 94.9 seconds on this 14-in, the middle 14-in machine, 81 seconds on the M4 Max, and 141 seconds on this one on the right. So I'm going to call it— that's the M4 and this is the M4 Pro.
So the M2 Max is faster at multi-core stuff than the M4 machine, but it has a lot more cores, so that's understandable. This is a multi-core operation. So if we weren't doing multi-core, if we were doing single-core things like JavaScript, for example, it's kind of hard to get a pure single-core operation these days.
But a lot of times when you're running a front-end process, like a web framework—think Angular, React, Vue, Svelte, Solid, things like that—that's going to be a one-core operation. For one-core operation, we have a test called Speedometer that does a pretty good job. I'm going to start that test right now.
What this is is just front-end JavaScript frameworks and doing to-do lists, adding items, subtracting items, showing graphs, and things like that. Wow, okay, the highest score that I've gotten so far—we haven't cracked 50 yet—but 49.1 on the M4 Max, 45.7 on the M4 Pro. We're going to call it that, and 47.4, which is higher than the M4 Pro on the M4 model.
But notice they're all very close, like within the margin of error. So we're talking about one core, and these have that one nice M4 core. They all have the same kind of core; they just have different amounts of them. And the M2 got a pretty decent score—40.8.
All right, for this next test, we don't need you. I'll be back. Yeah, this other test is a web-based test, but it's called Jetstream 2, and it's more than JavaScript. This has WASM, TypeScript, and it tests a bunch of other things like regular expressions and uglify—just a bunch of different web application libraries. Some of them are single-core, and some of them use multi-core libraries under the hood, which is why we're going to see some slightly different results between the M4 Pro and the M4.
So let's start off with the M2, which got a score of 379—pretty good for when it was back then. We got 497 for the M4 Pro, which is quite different than the M4—475. That's the test variants, but also it's a big enough score difference that I would say some of the multi-core operations in there had a role in it. And we got 504 for Jetstream 2 on the M4 Max—the highest score I've ever seen.
So it's lunchtime; I'm going to go eat lunch because this build is going to take forever. I remember it taking about 6 hours on my MacBook Air from 2015.
This is, of course, WebKit—basically Safari browser. It's not going to take that long on these machines, and it's going to take you a second. This is how we build it; I'm not going to do the M2 Max because I've already done that many times on the channel before. But this will tell you: if you have a really huge build, how long is it going to take you daily, and how many seconds is it going to help you save every day?
Eventually, you'll be able to add that up and take a vacation. All right, this is intense! I just had to take a look at this because the M4 Max is going nuts. Look at those fans—5,000 RPM and it's getting pretty toasty. Actually, it's not. It's 5° and holding, but the fans are going crazy.
This build goes in waves, and I've never seen the fans spin up this high—7,200 RPM, 7,900 RPM on fan 2. That's TG Pro, by the way, that's how I can see the temperatures and the fans. It is loud and warm in here, but I've got to say, the fans are doing an incredible job. They're much louder than in previous generations. I mean, I can tell that right now just from standing here and listening to it.
But this build is not going over 80°. So it looks like it's very important for Apple to keep these new machines cool and not exceed certain temperatures that they were allowed to exceed before. I wonder what the M4 MacBook Airs are going to look like because those don't have fans.
Now the M4 computer is not doing the same thing; the fan is low and the temperatures are pretty low. I'm also expecting that machine to come in last by far, and this makes me think that the M4 is significantly different than the M4 Pro and the M4 Max because they had to shove the M4 in the iPad first, which doesn't have a fan either. Interesting.
I want to see the results here, but I'm going to finish my sandwich. Okay, we got some results, folks! And yeah, that M4 took a long time. If you have to build this project a lot, you're going to get that vacation sooner if you get the M4 Max because we got 40 minutes on the M4 to build this project.
M4 Pro, actually not bad—24 minutes. This is the 14-core variety. I'll list the machines that I'm using down in the description, by the way, just below the like button. And the M4 Max—17 minutes! Wow, told you this was a big one.
Just as a side note, thanks to the members of the channel. I bought all these machines myself; nobody sent them to me, and members like you help support the channel. There's a join button down below too if you want to become a member; you're welcome to. You get extra videos, and you get to have badges by your name, and you can use special emojis.
Just a quick battery check while we're here. I started all these machines, the M4s, at 100%; we're down to 28% on the M4 Pro, we're down to 44% on the M4 Max, and 36% on the M4. The 16-in version does have a bigger battery, though.
Let's move on and do a little bit of torture. We're going to do something called the Mandelbrot test, and I've done this on the channel many times before. Those of you that are not new know about this. This is a Python program, so this time not compiled. We've done compiled now; it's interpreted, and this program is designed to utilize all the cores all at the same time—just max them all out, go really hard.
We're going to take a look at that and how that affects the performance. Of course, since this is a multi-core test, the machines with more cores are going to win, so you can kind of expect that. But what I'm curious about is to see how the M4 versus the M2 Max perform because they have the same number of cores.
Also, the M4 Pro has got 14 cores, and I think it should be the 12-core M2 Max, but let's see. I tested a lot of machines on this channel using this particular test—exact same test on Windows and on Macs—so you can go back and check those out, those times, if you're interested in that as well.
All right, here we go, and boom! They're all running excellent. Now while they're running, take a look at this: here is Activity Monitor and specifically the History chart for the CPUs. Look at that! Wow, you're done already? Come on!
By the way, I didn't write this program; this is from the Benchmark Game website. 16,000 is the parameter that they specify that I use. Anyway, you can see that green spike walking across the screen; that's when this program was running. So it wasn't that long—17 seconds on the M4 Pro, 14.6 seconds on the M4 Max, 29 seconds on the M4, and the M2 Max beats the M4—26.99.
I bet you can pick up an M2 Max machine right now for the same price that you can get the M4. Let's do a quick preview of machine learning here. I do have a video coming up with more detailed tests on machine learning specifically around the new M4 Max and how powerful it is, and also running a cluster of machines as well. So stay tuned for those videos; make sure you subscribe so you don't miss them.
But for now, I just want to see tokens per second using Llama. Llama is running; it's the latest version, and we're going to run Llama 3.2—write a 1,000-word story and go! That's the M4 Max. Of course, during the live stream, if you came to that the other day, this got 107.
Let's see if we can beat it, huh? 107—same score, 107 tokens per second, which is really nice. Of course, it is a small model—Llama 3.22 billion parameter model; it's not that big. And if you do this kind of thing, then you know that the larger the model in parameter count—13 billion, 35 billion, 70 billion—the more parameters, the slower the output.
Sometimes you run quantized versions of that, and you might have heard terms like F-16, which means 16 floating point, and then there's 8-4 INT, which are integer quantizations. Basically, quantization reduces the precision of model weights. So to go from F-16 to 4, for example, the model would become a lot smaller so it can fit on machines with less RAM.
It'll run faster, but you're sacrificing a little bit of performance and accuracy. It all depends on the model too. All right, here they are writing 1,000-word stories. This one is not going that fast. The M4, we got 68 tokens per second on the M4 Pro—quite a big difference from the M4 Max—and what has a huge effect on this is memory bandwidth.
The Max chips usually have twice the memory bandwidth than the Pro chips do. It's not exactly like that, but it's close. So we got 68 here; we've got 39 tokens per second on the M4, 82 tokens per second on the M2 Max—that Max chip.
So the M4 chip is pretty good at developer tasks, basic developer tasks. The Pro chip will get most developers—I'd say probably 90% of developers—to where they need to be, and the Max chip is really good for those people that are going to be doing lots of data, lots of number crunching, that need a lot of throughput—especially machine learning people.
Not too long ago, I did a video specifically related to memory bandwidth on Macs. Check it out right over here, and I'll see you very soon.