webcodr

Ryzen vs Apple Silicon and why Zen 3 is not so bad as you may think

Apple‘s M1 is a very impressive piece of hardware. If you look at the benchmarks, like SPECperf or Cinebench, it‘s an absolute beast with a ridiculously low power consumption compared with a Ryzen 9 5950X.

Zen 3 is that inefficient?

10 W vs 50 W looks really bad for AMD, but the answer to this question is a little more complicated than this raw comparison of power usage.

Here are basic things to know:

  • Apple‘s M1 has four high power cores (Firestorm) and four high efficiency cores (Icestorm), this is called BIG.little architecture.
  • AMD’s Ryzen 9 5950X is a traditional x86 CPU with 16 cores and 32 Threads

So what? The benchmarks are about single core performance and Ryzen needs 40 W more. What a waste.

Right, but this says nothing about the efficiency of a single core.

I don‘t have a 5950X, but it‘s older brother, the 3950X. Same core count and about the same power usage. While monitoring with HWINFO64 on a single core run of Cinebench R20 the used core needs about 10 - 12 W.

Okay, but what happened to the other 38 watts?

As you may know AMD introduced a so called chiplet architecture with Ryzen 3000. A Ryzen 9 5950X has three dies. Two with 8 cores each and a separate chip that handles I/O like PCIe 4, DDR4 memory and the Infinity Fabric links to the CPU dies.

This I/O die consumes 15 to 17 W. It‘s a quite large chip produced in a 12 nm node by GlobalFoundries. That alone is a big reason for higher power usage compared to a modern node like TSMC‘s N7P (7 nm) used for the other two chiplets.

Why 12 nm? It‘s a compromise, since TSMCs 7 nm production capabilities are still somewhat limited and AMD takes a fairly large share of the capacity. All Zen 2 and 3 cores are produced in 7 nm, as are all modern AMD GPUs and of course there are other customers as well.

I hope AMD will improve the I/O die with Zen 4 (Ryzen 6000) dramatically. A modern process node and a bunch of better energy saving functions would do the trick.

There are still 21 W unaccounted for!

15 more cores are also unaccounted for. 21 divided by 15: about 1.4 W per core average. Seems a bit high, but possible. It depends on background tasks from the OS, the current Windows power plan etc.

To be fair, AMDs energy saving functions per core could a little bit better. They are not bad but there‘s room for improvements.

Process nodes

For a fair comparison we also have to include the process nodes. As mentioned above, Ryzen cores are manufactured in TSMC‘s 7 nm node. Apple is one step ahead in this category. The M1 and it‘s little brother A14 are already on 5 nm, again from TSMC.

This alone could account for up to 30% less power usage. There are no exact numbers available, but 20 to 30% would reasonable.

Core architecture

Current Zen 3 cores and Apple‘s Firestorm cores have fundamental differences in architecture. Modern x86 cores are small but clock quite high. On a single core load like Cinebench R23 you will get about 5 GHz from 5950X while a Firestorm core will only clock to 3.2 GHz.

Wait, what? Firestorm with 3.2 GHz is about as fast as Zen 3 with 5 GHz? That can‘t be true. Yet it is. According to the benchmarks a Firestorm core can execute about twice the instructions per cycle (IPC) as a Zen 3 core.

Firestorm is an extremely wide core design with many processing units. This is complemented by absurdly large L1 caches and re-order buffers. Much larger than on any x86 core ever.

This allows Apple to lower the clock speed significantly without compromising performance. Lower clock speeds also mean lower voltage and power usage.

But there‘s more: since the A13 (iPhone 11) Apple has quite many incremental power saving functions within every part their SoCs. They can lower the power usage to an absolute minimum or even turn parts of the silicon on and off quite fast, to be ready when the user needs more computational power and as fast they turn it off.

Modern x86 CPUs have similar functions but not quite as advanced or in as many parts as current Apple SoCs. Remember AMD‘s I/O which has basically the same power usage all the time.

Instruction set

x86 is old and can be quite cumbersome to use. Apple’s CPUs are based on ARM, a more modern instruction set. The same tasks can be achieved with fewer commands. This means faster computation and less power usage.

Under the hood current x86 CPUs have noting to do with their older cousins, but the instruction set is still the same. To be fair, over the years AMD and Intel added way more modern instructions like AVX for certain operations, but the core is still 40 years old.

Conclusion

Taken all this into account, a Zen 3 is not that bad compared to a Firestorm core. Of course, Apple has advantages over x86, but they are more about the process node, different architecture or the instruction set, than about the efficiency of the cores itself.

Well, and then there‘s Intel. The quite slowly awakening giant is not without hope, but they are way behind Apple and AMD. As long as their high performance CPUs are stuck on 14 nm, they are … to put it simply, fucked.

10 nm is on it‘s way, but there a still problems. The next release of desktop- and server-class CPUs will still be on 14 nm, but with a more efficient architecture backported from 10 nm CPUs. It will help, but they need a better process node and they need it asap.

10 nm could be ready for the big guys in 2021. At least new Xeons in 10 nm were announced. But the question is: how good is the node? Some of the already released 10 nm notebook CPUs are quite good, so Intel could be back in year or two.

They also plan to release Alder Lake in 2021, a BIG.little CPU design. That would be much appreciated on notebooks, but software has to be ready. Without a decent support of the operating systems, such a CPU will not work properly.

There‘s another problem: Intel‘s 7 nm node. According to YouTubers like AdoredTV or Moore‘s law is dead, 7 nm could be the 10 nm disaster all over again. Let‘s hope not.

Three competing CPU vendors on par would be amazing. Not just for pricing, but also computational power. Look what AMD did in the past three years. If someone said in 2016 that AMD would kick Intel‘s and nVidia‘s ass in four years, he would have been called a mad man. The same with Apple.

2020 is a shitty year, but the hardware? Awesome would be an understatement.

webcodr goes Netlify CMS

Until today posts on webcodr were published via a simple git-based workflow. If I wanted to create a new posts, I had to open the repository in Visual Studio Code and created a markdown file. After pushing the commit with the new file, a GitHub hook notified Netlify to pull the repo and build and publish the site.

It was quite simple and effective, but lacks comfort and does not work on iOS/iPadOS devices. After buying a new iPad Air and Magic Keyboard, I wanted a pragmatic way to write posts without my MacBook or PC.

I always wanted to try Netlify CMS, so this was my chance. The transition was really simple. I followed the guide for Hugo-based sites and adjusted the config to work with Hugo. That‘s it. Netlify CMS works just fine with the already existing markdown files. Even custom frontmatter fields within the markdown files are no problem, just set them up in the config file.

A little more configuration in the Netlify admin panel is necessary, but the guide explains everything very well. It was just a matter of 15 minutes to get everything going.

If you use a static site generator and want to have a little more comfort, Netlify CMS makes it really simple. They provide guides for every major player like Gatsby, Jekyll or Nuxt and of course Hugo. You don‘t even have to use GitHub. GitLab and BitBucket are supported as well. As are more complex workflows for more than one editor, custom authentication with OAuth or custom media libraries.

Editor

Netlify CMS supports markdown and has a basic, but decent editor with rich-text mode. But I wanted a little bit more, so I decided to write my posts in Ulysses — a specialized writing app with support of GitHub-flavored markdown, including syntax highlighting preview . It‘s available for macOS and iOS/iPadOS. All files and settings are synched via iCloud. So, once you have setup everything, you‘re good to go on any of your Apple devices.

Since Ulysses requires a subscription, I will use this to „force“ me writing more posts. 😁

I wrote this post entirely on my iPad Air and so far, I‘m quite happy with the new workflow. Of course I will not write every post this way. New posts with code examples will be easier to handle on a Mac or PC. (Hey Apple, how about IntelliJ on an iPad?)

btw: even as an enthusiast of mechanical keyboards I have to say, it‘s quite nice to type on a Magic Keyboard. I just have to get used to the smaller size. The Magic Keyboard for the 10.9“ iPad Air or 11“ iPad Pro is a compromise in size and some keys like tab, shift, backspace, enter and the umlaut keys (bracket keys on english keyboard layouts) are way smaller compared to normal-sized keyboards. It‘s bigger brother for the 13“ iPad Pro has a normal layout, but that monster of an iPad seems a bit excessive for my use case.

Kotest and JUnit with IntelliJ or: don’t frak up your toolchain upgrades

My team and I recently decided to use Kotlin for new features in our existing project. It was a great choice to implement a new authentication process and we’re now rewriting some older parts of the application from Java to Kotlin.

Actually I wanted to use Kotlin for a while now, but there were only minor tasks within the Java part of the project. That finally changed and we can focus to improve the Java backend drastiscally.

Part of this process was a library update. We decided to upgrade JUnit from 4 to 5. A big pain in the ass. I don’t think, I would do it again. JUnit 5 was also part of a bigger problem, even if it was actually PEBCAC.

Kotest and MockK features

If you already know about Kotest or want to know more about the problem I had, just skip to next headline. The Kotest introduction is a little bit longer.

As I dived more and more into Kotlin, I stumbled over a Kotest. A really neat testing framework for Kotlin. There’s nothing wrong with JUnit, but Kotest gives you way more awesome ways to structure your tests.

A little example:

package io.webcodr.demo

import io.webcodr.demo
import io.kotest.assertions.throwables.shouldThrow
import io.kotest.core.spec.style.FunSpec
import io.mockk.*

class UserServiceTest : FunSpec() {
    private val userRepository = mockk<UserRepository>()
    private val service = UserService(userRepository)
    private lateint var user: User

    init {
        beforeTest {
            user = User(1, "Jane", "Doe", "[email protected]")
        }

        afterTest {
            clearAllMocks()
        }

        context("getUser()") {
            fun verifyRepoCalls() {
                verify {
                    userRepository.findById(1)
                }

                confirmVerified(userRepository)
            }

            test("should succeed") {
                every {
                    userRepository.findById(1)
                } returns user

                service.getUser(1).shouldBe(user)

                verifyRepoCalls()
            }

            test("should fail") {
                every {
                    userRepository.findById(1)
                } throws UserNotFoundException()

                shouldThrow<UserNotFoundException> {
                    service.getUser(1)
                }

                verifyRepoCalls()
            }
        }
    }
}

Kotest offers serveral different styles to write tests. I chose the FunSpec style for this example. You could also use a BDD-like or Jasmine-like style, if you want to.

It’s much more intuitive to nest tests with Kotest. To be fair, JUnit 5 allows you to use @Nested with an inner class to acomplish nesting as well, but it’s not as intuitive and harder to read than trailing lambdas.

Assertions are also a little easier to write. Kotest has over 100 different matchers. You can use them as extensions functions, like in the example above or alternatively as infix functions, for example service.getUser(1) shouldBe user. It’s also quite simple to write custom matchers.

There are many more features like soft assertions, tagging, easy temporay file creation or handling for non-deterministic test cases.

For mocking we decided to use MockK, since it’s way more intuitive to use than Mockito with Kotlin. Don’t get me wrong, Mockito is a great library, but it has one flaw: the when method. when is a keyword in Kotlin and to use it with Mockito, you need to write it in backticks. That’s quite ugly and not intuitive at all.

So, in summary, Kotest offers a bunch of pretty neat features and is very intuitive to use. Of course, JUnit can achieve much of this as well, it’s just not that shiny and little harder to read.

The actual problem or: why the frak is there IntelliJ in the title?

If you migrate an old codebase to Kotlin and want to use Kotest, you will have no choice and have to use JUnit and Kotest in coexistence.

That shouldn’t be a problem, since Kotest uses the JUnit 5 Jupiter engine under the hood. But …

As I wrote a new service in Kotlin and some tests with Kotest, I could not start the JUnit tests anymore. As soon as the maven depedency of Kotest was present, IntelliJ didn’t recognize JUnit tests and used the Kotest files only. With mvn test (Maven Surefire) everything worked fine.

I tried several things and was ready to give up. Search engines didn’t find anything about this problem. Nothing on GitHub, nothing on Stack Overflow.

I hate to give up, so I decided to create a small demo project to open a GitHub issue. Well, that didn’t work out as intended, since the discovery of JUnit tests in the demo project worked fine. The IntelliJ JUnit runner did what it was supposed to: run JUnit and Kotest.

Well, frak. There must be some kind configuration problem with my real project. I already looked at the IntelliJ runner config, Maven files etc. — nothing worked.

I compared the Maven files from both codebases and there was one difference: my real project did not include the JUnit Jupiter Engine depedency. Bingo. I added to the Maven file and guess what? It worked like a charm.

What an embarassment. As we upgraded from JUnit 4 to 5, we forgot to add the new depedency for the engine. I don’t know why the tests worked at all, but it seems the engine is not really necessary for all cases. But it can screw up test discovery quite well, if you forget it.

The dependency configuration in POM file should like this:

<dependency>
    <groupId>org.junit.jupiter</groupId>
    <artifactId>junit-jupiter-api</artifactId>
    <version>${junit.version}</version>
    <scope>test</scope>
</dependency>
<dependency>
    <groupId>org.junit.jupiter</groupId>
    <artifactId>junit-jupiter-engine</artifactId>
    <version>${junit.version}</version>
    <scope>test</scope>
</dependency>

Well, that was a long post, but IMO it was necessary to show how this problem came to be and even if you have no trouble at all, perhaps you’ll consider to use Kotest. It’s awesome!

Gradle

I’m not sure, but this issue could happen with Gradle as well, when you are migrating to JUnit 5.

macOS Catalina EDID Override AKA HDMI color fix

HDMI connections from your Mac to monitor can be a pain in the ass. There is a chance that macOS will detect your monitor as a TV and set the color space to YCbCr. You will get wrong colors and sometimes blurry fonts.

If you’re having this problem, like me, you know the fix: a patched EDID created with this little Ruby script.

The installation of this EDID override could be tedious since the release of El Capitan, as SIP won’t let you access the necessary system files. Just disable it in recovery mode, copy the file and enable it again. Sucks, but works just fine.

Now, Catalina is out for a few hours and has a new way to annoy people who need EDID overrides. All system-related directories and files are read-only, regardless of the status of SIP.

Fortunately Apple was not crazy enough to disable the write access completely.

Help is on the way, ETA 0 seconds!

  1. Patch your EDID
  2. Boot into recovery mode with CMD+R
  3. Login with your user
  4. Open Disk Utility, select your volume (in most cases Macintosh HD) and mount it with your password (yes, again)
  5. Close the Disk Utility and open a Terminal window
  6. Copy the directory with the patched EDID to /Volumes/$VOLUME_NAME/System/Library/Displays/Contents/Resources/Overrides
  7. Reboot and enjoy the right colors again

Here’s an example of the shell commands:

cd /Volumes/Macintosh\ HD/System/Library/Displays/Contents/Resources/Overrides
cp -rf /Volumes/Macintosh\ HD/Users/webcodr/DisplayVendorID-5a63 .

Don’t forget to use the correct volume and user names!

Dear Apple

Just add a simple solution to select the HDMI color space. A simple shell commando with sudo would suffice or at least let us use an override directory within the user library as it was possible many years ago. It just sucks to do this after every macOS upgrade and every time you improve system security, it gets harder.

Please, don’t forget us powers users …

Hello, Dark Mode

Dark mode for Android and iOS? Hold my beer …

It’s quite simple to implement. Every modern browser can evaluate media queries in JavaScript with window.matchMedia() and supports CSS variables.

I added the following to my application JavaScript file:

const preferColorSchemeResult
  = window.matchMedia('(prefers-color-scheme: dark)')

if (preferColorSchemeResult && preferColorSchemeResult.matches === true) {
  document.documentElement.setAttribute('data-theme', 'dark')
} else {
  document.documentElement.setAttribute('data-theme', 'light')
}

The script will set the data attribute theme on the document element (html) with the possible values dark or light depending on the result of the media query.

There’s no need for a polyfill, even IE 10 supports window.matchMedia()

Stylesheet changes is even simpler, since I already had introduced SCSS color variables a while ago. I just had to replace them with CSS variables.

// colors
$c_white: #fff;
$c_dark-grey: #4A4A4A;

:root {
  --container-background-color: #{$c_white};
  ...
}

[data-theme="dark"] {
  --container-background-color: #{darken($c_dark-grey, 20%)};
  ...
}

That’s basically it. If you use SCSS, please take notice to use interpolations to map the SCSS variables to CSS variables. This change in SassScript expressions was necessary to provide full compatibility with plain CSS.

Since the theme selection is fully automated, I will provide a toggle possibiliry in a future release for those of you who prefer the light mode. This can be easily achieved with a flag in local storage and some minor changes in the JavaScript part.