Tuesday, May 28, 2024
HomeTechnologyApple's new M1 chip-equipped MacBook is a machine learning beast machine !

Apple’s new M1 chip-equipped MacBook is a machine learning beast machine [Part 1]!

Table of contents

  • Speed ​​tests of the (almost) top-spec Intel-based 16-inch MacBook Pro, the new Apple silicon MacBook Air, and the 13-inch MacBook Pro
  • why test/compare
  • Mac specs
  • test
  • Experiment 1: Final Cut Pro video export
  • Experiment 2: Training a machine learning model using CreateML

Speed ​​tests of the (almost) top-spec Intel-based 16-inch MacBook Pro, the new Apple silicon MacBook Air, and the 13-inch MacBook Pro

I watched the M1 chip keynote, looked at the graphs, looked at the battery life, and was instantly blown away. And I was drawn to it. Then I started thinking. How could one of these new M1 chip-equipped MacBooks fit into my life?

Of course, I didn’t need a MacBook with an M1 chip. But I kept wondering what story I could tell myself to justify buying another computer. Then I came up with an excuse to buy. My 16-inch MacBook Pro is too heavy to carry around all the time. Oh, that’s fine. This 2.0kg aluminum high-spec machine is too heavy to carry around.

Wait a minute… 2.0kg is 4.4lbs?

Was it just that heavy?

That’s right.

Wow. It’s not that heavy.

Please don’t get in the way of telling a good, truthful story.

I ended up making an excuse to buy a MacBook with an M1 chip. Here’s why I ordered a shiny new one (or two). My 16-inch MacBook is too heavy to carry to a cafe to code and write, edit videos, and check email sporadically.

And Apple says the new M1 chip is 11x, 15x, 12x, or even 3x faster for different things. These are suggestive numbers, but I haven’t measured them in the past.

All I care about is whether I can do what I need to do quickly.

The last word in the previous sentence is the most important. When it comes to PC specs, there are things that are no longer negotiable. Speed ​​is part of me now. Since moving from hard drives to solid state drives, there’s no going back.

I purchased a 16-inch MacBook in February 2020. I had just completed a big project and had plenty of cash, so I decided to buy myself a workstation that promised my future. I edit videos every day and hate processing delays, so I wanted to choose a product that would be the best partner I could buy, so basically all the specs were maxed out except for storage (see the table below for the specs) reference).

Luckily, I had a friend (thanks Joey) who could use the Apple Store Clerk’s discount to purchase the Beast Machine.

Anyway, I already talked about my number one criterion, speed. Consider other requests as well.

  • speed. If you don’t go fast, you’ll get lost.
  • cost. Cost is also a big factor in demanding from a PC, but I don’t want to pay for a higher spec machine or just a higher quality computer. The PC is my first tool. It’s what you use to make art, make money, learn, and connect with the world.
  • portability. I hate sitting in the office all day. Can you take it to a cafe or library for hours without looking for an outlet? Consider portability based on a combination of battery life and weight.

why test/compare

Why not test it?

Actually, I’m a geek. Also an Apple fan. And I wanted to see how my big, top-of-the-line 16-inch MacBook Pro compares fairly to the one with the new M1 chip.

What’s more, the excitement I felt about a new computing device, the M1 chip, wasn’t there since the introduction of the original iPhone.

Other reasons included carrying a light laptop and wanting tax incentives (you can claim the incentives if you buy another one before the end of the year).

Mac specs

Whenever I buy a new machine, I upgrade the RAM and storage a step or two from the baseline.

512GB of storage and 16GB of RAM seems like the bare minimum for me these days (seriously, I doubt anyone actually runs a 128GB MacBook).

So I upgraded the RAM on both M1 MacBooks from 8GB to 16GB, and the storage on the 13-inch Pro from 256GB to 512GB.

A 16″ MacBook is my current machine and it worked fine until I ran the following tests.

(The table below summarizes the specifications of the above three MacBooks.)

M1 chip MacBook Air M1 chip 13-inch MacBook Pro Intel-based 16-inch MacBook Pro
CPU 8-core M1 8-core M1 2.4GHz 8-core Intel Core i9
GPUs 7-core M1 8-core M1 AMD Radeon Pro 5500M with 8GB GDDR6 memory
Neural engine 16-core M1 16-core M1 none
memory (RAM) 16GB 16GB 64GB
storage 256GB 512GB 2TB
weight 1.29kg 1.4kg 2.0kg
Maximum continuous battery life (estimate) ~15 hours ~20 hours ~11 hours
OS macOS Big Sur 11.0.1 macOS Big Sur 11.0.1 macOS Big Sur 11.0.1
* (actual) price $1,899 AUD ($1,430 USD) $2,599 AUD ($1,950 USD) $6,649 AUD ($5,000 USD)
** (baseline) price $1,899 AUD ($1,430 USD) $2,299 AUD ($1,730 USD) $6,049 AUD ($4,550 USD)

(Table: Specs table of MacBook Pros tested) Specs of each Mac tested. Note: I usually upgrade RAM and storage, even though there are cheaper configurations for each model.

  • *(Real) price is the price I paid for each model. Note: I actually paid $5,500 AUD for the MacBook Pro 16-inch because I have a friend who works at Apple who gave me his employee discount (thanks Joey).
  • **(Baseline) price is the price after upgrading all processing components except storage (e.g. 8GB → 16GB RAM for M1 model, 2.3GHz → 2.4GHz for Intel model). (Excluding storage because that upgrade is the most expensive.)
(*Translation Note 1) “AUD” in the table above means Austrian dollars and “USD” means US dollars .


Apple’s chart I saw in the keynote was impressive. And the GeekBench score was even more impressive. But these are just numbers for me. I wanted to see how a MacBook with an M1 chip would perform in what I actually do on a daily basis. These daily tasks are:

  • making document. I’m assuming every MacBook is good at writing documents.
  • web browsing. Same as above, assuming no weak machines.
  • video editing. It was one of the main uses I was interested in, and one of the main reasons I bought the 16-inch MacBook Pro with a dedicated GPU.
  • code creation. I don’t really need a text editor, but Xcode is getting pretty heavy these days.
  • Training machine learning models. I write a lot of machine learning code. I don’t expect to be able to train state-of-the-art models on a laptop, but at least I’d like to be able to play around and experiment.
(*Translation Note 2) Geekbench is a cross-platform benchmark software . The supported OS is windows, Mac, Linux, Android, iOS. Although it is paid software, the tryout mode that automatically posts the score to the sharing site is free. You can search the score of the M1 chip-equipped MacBook from the Geekbench Browser , which allows you to view the uploaded score .

After reviewing the use of my PC as described above, I devised three tests.

  1. Video export using Final Cut Pro. Earlier this year I made a fairly lengthy video ( on Machine Learning Roadmap 2020 ), which was over 2 hours and 37 minutes. So I thought it would be cool to see how long each machine takes to write out this video.
  2. Training machine learning models with CreateML. CreateML is a black box machine learning model creation function developed by Apple. I don’t have a large Xcode file ready to go, but I decided to take a look at how CreateML handles machine learning model training on the new Apple silicon .
  3. Running TensorFlow native code using tensorflow_macos . The test I was most looking forward to. Apple and the TensorFlow publisherspublished a blog post stating that the new TensorFlow for macOS fork dramatically speeds up training machine learning models on the new M1 chip. Is this claim true?
The “about” on the tensorflow_macos GitHub page states that ” TensorFlow for macOS 11.0+ is accelerated using Apple’s MLCompute framework .”

Why not test more extensively?

That’s enough for me. I have other things to do.

So it’s time to show off the test results. The best results in each experiment are highlighted in bold.

Experiment 1: Final Cut Pro video export

This time, we gave all machines plenty of time to pre-render the RAW video. So when the export button is clicked, all machines should be showing relatively the same page.

Experiment details:

  • Video Length: Over 2 hours 37 minutes
  • Export file size: 26.6GB (estimate), ~6.5GB (actual size)
M1 chip MacBook Air M1 chip 13-inch MacBook Pro Intel-based 16-inch MacBook Pro
writing time 38 minutes 10 seconds 32 minutes 38 seconds 21 minutes 23 seconds
Battery level at start 89% 92% 87%
End battery level 78% 83% 51%
Battery level change -11% -9% -36%

Not surprisingly, it was the fastest exported on a 16-inch MacBook Pro. The most likely reason is that it has a dedicated 8GB GPU and 64GB of RAM.

However, at the cost of using a dedicated GPU, the battery consumption and fan rotation speed were amazing (in the video recording the test, you can see the fan running out of control like a jet plane).

During the video export, the MacBook Air with M1 chip and the MacBook Pro 13″ remained completely silent (obviously the MacBook Air doesn’t have a fan, but the MacBook Pro 13″ fan never turned on. ).

Experiment 2: Training a machine learning model using CreateML

I’ve never actually used a machine learning model trained with CreateML before. But we decided to take a look at how one of Apple’s custom features could take advantage of the company’s new silicon-powered PCs.

In this test, the CreateML settings for each MacBook were as follows.

  • Challenge: Multi-class image classification.
  • Model: Image classification using CreateML (Apple can’t tell me what architecture they use, but I presume it’s ResNet ).
  • Data size: 7500 training images, 2500 test images
  • Maximum number of iterations: 25
  • Data extension: flip, rotate

Experiment 2 setup: training an image classification machine learning model using CreateML, 7500 training images, 2500 test images, 25 max iterations (model seems to converge in 10), flip/rotate data augmentation.

M1 chip MacBook Air M1 chip 13-inch MacBook Pro Intel-based 16-inch MacBook Pro
model training time 11 minutes 30 seconds 15 minutes 30 seconds 43 minutes, 10 minutes after restart (incomplete)
Battery level at start 72% 75% 31%
End battery level 68% 69% 0%*
Battery level change -Four% -7% -31%

(Table: Experimental results using CreateML) Training results of an image classification model using CreateML. Surprisingly, the MacBook Air performed the fastest compared to the 13-inch MacBook Pro’s 8-core GPU M1, even though it has no fan and a 7-core GPU M1.

  • *MacBook Pro 16″ died before the end of the test.

The most surprising result of all the tests was that the MacBook Air with the M1 chip won by a clear margin in both training time and battery life. Even though this model has one less fan and GPU than the 13-inch MacBook Pro (the MacBook Air I used had an M1 chip with an 8-core CPU and a 7-core GPU).

These results also reveal that the CreateML function may have been optimized for the M1 chip. Despite having eight CPU cores and a dedicated GPU, the 16-inch Intel-chipped MacBook Pro ran out of power before I finished the experiment.



Please enter your comment!
Please enter your name here

Recent Posts

Most Popular

Recent Comments