This is the laptop that always makes me feel kind of stupid cuz normally, since I have [music] this YouTube channel, I have the Max version. This is
the M4 Max and now I have the M5 Max, which is supposed to be the ridiculous one. And then I go back and test the Pro chip, open the same apps, do the
same work, push it through the same nonsense, and I start having some very uncomfortable thoughts.
So, I decided to test everything, not just the synthetic benchmarks, real builds, real SSD transfers, real AI models side by side. And what I found is
that some of these results are embarrassingly close. Oh, what? Between these two. Not that one. And some of them are not even in the same universe.
Okay. Let's start with what you can actually see. By now, you've already seen all the specs of the
whole M5 family, I'm sure. Plus, you can get this from the Apple website. But, here's the speed run of all the important items that affect developers.
14-in M5 base model. The M5s don't come any bigger. M5 Pro and M5 Max are in 16-in because that's how I prefer it. I have more real estate for my
code. Personal preference, you can argue me in the comments. Since I'm looking at performance, the key
numbers are 10 CPU cores on the M5 versus 18 CPU cores on the Pro and the Max. These are just the models I have. There's other variations. One fan on
the M5, two fans on the Pro and the Max. [music] Now, the memory bandwidth, that roughly doubles at each step. 153 GB per second, 307 GB per second,
and 614 GB per second. Although, those are just theoretical numbers and I'll show you the real
numbers later. Two things to flag before we start testing. The Pro and the Max get Thunderbolt 5, three of those. The M5, Thunderbolt 4, two times
slower. And both the Pro and the Max support high power mode. And that lets you push the fans harder for sustained workloads. The M5, well, that one
doesn't have that option. You only get the dial it down low power mode option, not push it harder like
the M5 Pro and M5 Max. Keep it in mind, it's going to come back later. >> [music] >> Every time you clone a repository, install dependencies, or open
a project, or download an AI model, that's your SSD it's hitting. And there are actually three different speed stories here depending on what you're
doing. Oh, what? SSD speeds are much faster on the M5 Pro and the Max since Apple adopted PCIe Gen 5.
And disk speed depend on the size of the drive. Higher capacity models have more NAND chips for parallel access. And they also have larger SLC caches.
And you can see that here. My M5 Max is 4 TB, M5 Pro has 2 TB, and this one has 512 GB. These are some crazy numbers. And even the slowest one here
should be enough for most use cases. For reference, the M4 at 512 gigs did about 2600 MB per second
read. The M5 Max at 4 TB, 7200. Now, the M5 Pro and Max, 13,200, 13,500. And for write, 15,400 and 17,300. That's roughly double what the M4 Max did.
That's PCIe Gen 5 at work. M5 is no slouch either. That's like two and a half times faster than M4. But, is that what real file operations look like?
I copied a 4.5 GB LO M file from one folder to another on the internal SSD. The M5, 2.1 GB per
second. The M5 Pro, 8.9. And the M5 Max, 10.4. Real numbers. And that confirms the synthetic story. PCIe Gen 5 on the Pro and Max is real and you can
see it in raw sequential copies. Then I copied the entire WebKit source tree, over half a million files, 43 GB in tiny source code files. Can you
imagine how long it took them to write that? And they didn't byte code that. M5, 260 MB per second. M5
Pro, 276. M5 Max, 306. A little different, huh? But, only an 18% spread between the M5 and the M5 Max. So, for internal many file operations, all
three machines handle metadata roughly about the same speed. The internal SSD controller isn't the differentiator here. So, how fast can you move data
to an external drive? This is about the Thunderbolt bus, not the internal SSD. I plugged identical
Thunderbolt external SSDs into all three machines and copied the same 5 GB model files. The M5 on Thunderbolt 4, about 1.4 GB per second write, 3.9
read. And now we see a big difference. The M5 Pro on Thunderbolt 5, 3.4 write, 9.6 read. And the M5 Max, nearly 6 GB per second write and almost 11
read. Clean scaling. Thunderbolt 4 caps the M5 and the Max's faster internal SSD controller gives it an
edge over the Pro. A slight edge, even though they're both on Thunderbolt 5. Then I did the same thing with the half a million small files. The M5 and
the M5 Pro, essentially identical. 67 to 69 MB per second write, nearly 10 minutes each. Thunderbolt 5 didn't help at all here. The bottleneck isn't
the bus bandwidth, it's IOPS and file system metadata. The M5 Max, five times faster at 336 MB per
second. That's the different storage controller doing its job. And that's what real dev workflows actually look like. Cloning a repo, copying node
modules. Node modules. Oh, boy. So, when somebody says my SSD benchmark is at 7 gigs a second, ask them what they're copying. AI coding tools are
fast. The problem is the overhead they create. I still end up testing different models, weighing answers,
and turning scattered output into a plan myself. Verden is the first tool I've used that manages the workflow around the code, not just the code
itself. Take multi-plan mode. Instead of picking one model and hoping it was the right call, Verden has three leading models weigh in together. So,
I've got a .NET issue tracker I maintain and I asked Verden to plan out a new dashboard feature for me.
Claude came back with a detailed production-ready approach that keeps the new stuff isolated and safe. Gemini went the other direction, more of a fast
prototype that just modifies what's already there. Verden takes both and gives me one consolidated plan that I can refine or just let it run. From
there, I pick the model and effort level for implementation. And with the help of the right skill,
ASP.NET Core in this case, we're off to the races. There's also a multi-mode code review built in. Verden's own benchmarks have it beating Codex
Connector on precision and recall using 40% less tokens. And at any point, I can pop into VS Code to keep an eye on things because I've still got some
control issues I'm working through. But, I'm getting better. This one's for people who are already
building with AI, but are tired of managing the workflow around it. It's less about piling more AI and more about cutting the busy work AI creates in
the first place. You can check out Verden at the link in the description. Now, back to the video. Now, let's get into the dev tests. And I'm going to
start with the ones where these machines are surprisingly close before we get to what they're not
close at. >> [music] >> Speedometer and JetStream, the classic browser benchmarks. And as expected, these are close. The M5 already has a strong
single-core CPU. And since these are all the same, pretty much the same core, and these tests are mostly single-threaded JavaScript work, the gap
between the M5 and the M5 Max here is not that huge. 61.4 on the M5, 60.2 on the M5 Pro, and 61.7 on the M5
Max. And there's your JetStream results, too. Slightly more detailed as a few more different workloads. And that could explain potentially the slight
difference we see in the test results. But, that is the point. Everyday browser responsiveness, web app snappiness, you're not going to feel a big
difference between these machines. So, this is for the dev testing perspective and the end user
perspective. The M5 is fast, that is not where the money goes. So, where does it start to matter? Let's add more cores to the equation. This is a
Python Mandelbrot set computation at 16,000 pixels using all cores via multiprocessing. Pure number crunching. No IO and no compiler complexity, just
how fast can you use all your cores? Look at all those cores being used. And no matter what machine
we're looking at, the cores are all going to be used. So, obviously, the machine with more cores is going to win. M5, 29.4 seconds. M5 Pro, 12.2
seconds. Big difference. And M5 Max, 11.7. A little bit faster. But, the Pro and the Max are nearly the same. Same 18-core CPU, just half a second
apart. So, right away, my Max is looking at the Pro and wondering what exactly it's paying for. But, the M5
at 10 cores is two and a half times slower. For context, the M4 was 29 seconds, too. So, the M5 base is basically the same speed as last year for
multi-core work. The improvement came from adding more cores to the Pro and the Max, not from making the same cores faster. All right, so pure compute
scales linearly with cores. But, real builds involve compilers, linkers, IO, and memory pressure. Does
the gap hold? In comes Xcode benchmark. 70-plus CocoaPods mixed Swift and Objective-C, C, and C++. This has kind of become the standard iOS dev build
benchmark. M5, 154 seconds for this. By the way, the lower the better, obviously. Builds it faster. M5 Pro with 18 cores, 89 seconds for that one. And
the M5 Max, 85 seconds. Well, here, the M5 is nearly twice as slow as the Pro, but these two are
pretty close. And here's a fun fact, the base M5 at 154 seconds is actually slower than the 3-year-old M2 Max, which did this in 127 seconds. More
cores beat a faster clock every time. Now, I should note, my M5 Pro is the full 18-core model. Max Tech tested the 15-core M5 Pro and got 99 seconds.
And Max's M5 Max, wow, I just said Max so many times. They got roughly the same as mine. So, even
within the Pro tier, those three missing cores cost you about 10 seconds in this particular build. Something to keep in mind when configuring. But,
there's another story hiding in this chart. First time I ran this, I ran it immediately after a long WebKit build, which I'll go into in a moment,
when the machines were still hot. Bad Alex. Bad. But, it was interesting because the M5 has only one fan.
And with its single fan and no high power mode, it went from 154 seconds cold to 246 seconds hot. That's 60% slower, just from thermals. The M5 Pro
lost 37% and the M5 Max barely moved. Remember when I said keep the power modes in mind? This is why. The M5 doesn't just have fewer cores, it can't
sustain what it has. But, Xcode benchmark is a medium-size project. What happens when we throw
something truly massive at these machines? A full clean WebKit release build. Millions of lines of C++, hundreds of compilation units. M5 Max, 15.3
minutes. That's probably the fastest I've ever seen this test done. M5 Pro, 16.1 minutes. Less than a minute apart. Again, I'm starting to see a
pattern here that my wallet doesn't love. Now, M5, 56 [music] and 1/2 minutes. That's almost an hour versus
15 minutes. 3.7 times slower. This is the single most dramatic result in this entire comparison. If you compile large projects regularly, the base M5
is not in the same league. The 10 cores, the single fan, the thermal constraints, they all compound over a sustained build like this. Okay, but not
everybody builds WebKit. What about the web developers in the audience? That's what I do, actually.
Victor Savkin, one of the founders of NX, created this large mono repo test, which I like to use sometimes. It's pretty cool. This is a 253 package
TypeScript mono repo with over 5,000 files. And here's where things get a little bit uncomfortable again for M5 Max owners like me. Non-batched cold
builds, M5 took 48 seconds, M5 Max took 42. That's a 15% difference. Not huge considering the price
difference between these two. And in batch mode, basically identical across all three, about 13 seconds each. TypeScript compilation trans polation is
mostly single-threaded. It doesn't scale with core counts like C++ does. If you're a web dev working with TypeScript and node, the base M5 is really
fine. And that's an uncomfortable thought when you spend three times as much on the Max. Let's try
another ecosystem and we'll see why memory matters. 100,000 generated C# classes and namespaces. This test is interesting because the previous M4
results showed a 25-second gap just from doubling the RAM, 16 gigs versus 32 gigs with the same CPU. So, it's not just about cores. M5 with 16 gigs,
84 seconds for this build. M5 Pro with 64 gigs, 55 seconds. And M5 Max with 128 gigs, 51 seconds.
Obviously, core count is the primary driver here, 84 versus 51 seconds. But, the Pro is only 4 seconds slower than the Max despite having the same
cores. That is memory bandwidth showing up in the build test, but just barely, 4 seconds. Just a few thousand dollars more. I keep writing these
numbers down feeling worse about my choices. Now, the M5 with 16 gigs, faster than last year's M4 Air at 94
seconds, is still firmly in last place among the M5 family. So, the Pro and the Max have been basically tied this entire time, but there's a whole
other dimension to test. And this is where the Max finally gets to justify its existence. Maybe. Let's see. >> [music] >> Before I run any AI models,
I want to show you something because local LLM performance depends on exactly two things, and we can
measure them both directly. The [music] first is memory bandwidth, and that's what I'm doing right now. Boom. Ooh. That's how fast the machine can
read data from memory. This directly determines the token generation speed, tokens per second, because for every single token the model generates, it
has to read the entire model's weights from memory. Stream Triad, the gold standard for practical
memory bandwidth and has been around for years. M5, 115 GB/s, M5 Pro, 251, and M5 Max, 350. Now, Apple's spec sheet says 153, 307, and 614. Our
measured numbers are lower, and that's normal. Every vendor spec is a theoretical ceiling, but the ratio still holds, roughly 1x, 2.1x, and 3x.
Remember that ratio because we're about to see it again. This is the same MLX framework that runs local AI
models doing matrix multiplications on the GPU. M5 got 8.1 teraflops, M5 Pro 19.7, and M5 Max 42.4. At Bfloat16, 15, 30, and 57 teraflops. That's
almost perfect scaling with GPU core counts, 10, 20, and 40. Yes, the GPU core counts. We haven't talked about that yet. This is where they come out
to play. Now, for the M5 Max, Apple says 614 GB/s >> [music] >> and 40 GPU cores. But, what can it
actually sustain? Now, we know. So, while bandwidth scales 1 to 2.2 to 3, compute scales from 1 to 2 to 4. Let's take a look to see if the AI
performance matches. Gemma 4 E4B. Yeah, it's new, so what? MLX can run it. Oh, MLX is the framework that runs LLMs on Apple Silicon. It's optimized.
It's kind of like llama.cpp, but optimized for Apple Silicon. This is an 8.2 GB model. Fits on all three
machines. Token generation on M5 at 25 tokens per second, M5 Pro at 51, and M5 Max, 87. There's a difference. That's 1x, 2.1x, and 3.5x. The bandwidth
ratio is almost exactly right. The raw hardware benchmarks we just ran directly predict real AI performance. And finally, we can say that the Max is
meaningfully faster than the Pro here. Not 4 seconds faster, not half a second faster, 70% faster.
This is what bandwidth buys you. After spending all this time watching the Pro keeping up, this is the first test where I don't feel completely stupid
for buying the Max. Now, all three machines can run this model, but the experience is different. 1 2 3. Just for kicks, I ran Gemma 4 in LM Studio, so
I have a comparison point with some of the ones I've been testing. 27 tokens per second on the M5,
98 tokens per second on the M5 Pro, and 166 on the M5 Max. But, what happens when we step up to a bigger model? >> [music] >> Quant 3.5 27B at 4-bit
quantization, 15 GB. The M5 Pro runs it at 17.6 tokens per second, the M5 Max at 25.1. Both usable. The M5? It crashed. A 15 GB model on a machine
with 16 GB of total memory, uh there's nothing left for the operating system. It's not slow, it's
broken. This is the memory wall. It's not about speed anymore, it's about whether the model fits at all. >> [music] >> And then there's this. Quant
3.5 122B, a mixture of experts model at 65 GB. The M5 Pro maxes out at 64 GB of RAM, and this model is 65 gigs before the OS even loads. It literally
cannot exist on this machine. The M5 Max with 128 gigs, loads it fine, generates at 23 tokens per
second. And here's the wild part. That's almost the same speed as the dense 27 billion parameter model because only 10 billion parameters are active
per token. Mixture of expert models or MoE, like this 122 billion parameter model, give you a much smarter model at the same inference speed. But,
only if you have memory to hold it. This is the why you buy the Max demo. Not because it's a little
faster, because entire categories of models simply don't exist in any other machine. Okay, I feel slightly less stupid now. So, do you remember how
the M5 throttled 60% on compilation after getting hot? Well, I ran 20 back-to-back AI completions to see if the same thing happens during inference.
And the M5 gets 22.8 tokens per second, plus or minus, and that's first to last run. The M5 Pro, 48.4
tokens per second, plus or minus, and the M5 Max, 82.4 tokens per second, plus or minus. All three are perfectly flat. There's no throttling on any of
these machines, including the single-fan M5. Why, you might ask? Well, because LLM inference in this case is memory bandwidth bound. The tests I'm
doing are not compute bound. It doesn't push the CPU or GPU hard enough to generate the kind of heat
that causes throttling. Whereas compilation that we did earlier, it hammers all the cores at maximum power for minutes. Inference just reads from
memory. So, the thermal story is workload dependent. The M5 throttles on builds, it doesn't throttle on AI, and that's a nuance that a lot of reviews
miss. So, for browsing, writing code, TypeScript builds, general productivity, the base M5 is
embarrassingly close to both. It genuinely makes me question every dollar I spent on the Max. At one point, when these came out in October, I
considered switching from the M4 Max to using just this, and then doing heavier workloads on my Mac Studios, like video editing and AI. For Xcode
builds, large compilations, anything that hits the cores with more than just a few minutes, the M5 falls off a
cliff a little bit, though, as we've seen. But, the Pro handles all that as well as the Max. For local AI, though, that's where the lineup finally
makes sense for a three-tier product. The M5 can run small models, the M5 Pro [music] can run serious ones, and the M5 Max can run models that
literally don't fit anywhere else, and do it really fast. But, the M5 Pro is the real developer sweet spot.
Professional iOS and macOS developers, full-stack developers with Docker heavy workflows, and serious AI enthusiasts who want to run 70 billion
parameter models. Really? Quantized. Quantized. Okay. So, the M5 Pro is the one that makes me feel the most stupid because it does almost everything
the Max does for significantly less money. Honestly, I think most people that are software developers that
are professionals should just buy the M5 Pro and not feel bad about it. I'm the one that's an idiot that bought all these MacBooks. So, I should feel
bad. Anyway, if you want to see my first look at the M5 Max, a little bit of a deeper deep dive, check out this video right here. Thanks [music] for
watching, and I'll see you next time.
[music]