X E-LEAP

How far we have come in a generation. Snapdragon X Elite laptop vs Windows Dev Kit 2023 vs Windows on Arm Virtual Machine running on MacBook Pro in developer related tests.

Use my coupon ZISKIND10

🛒 Gear Links 🛒

🎥 Related Videos 🎥

— — — — — — — — —

❤️ SUBSCRIBE TO MY YOUTUBE CHANNEL 📺

Join this channel to get access to perks:

— — — — — — — — —

📱LET'S CONNECT ON SOCIAL MEDIA

— — — — — — — — —

#Windows #SnapdragonXElite #armprocessors

📖 Chapters 📖

0:00 - Introduction
0:33 - Comparing Snapdragon Generations
0:53 - Windows on ARM options
1:12 - Testing Setup for Windows 11
1:59 - Python Performance Test Introduction
2:25 - Running the Mandelbrot Python Test
3:43 - Adjusting Virtual Machine Settings
4:09 - Surface Performance in Python Test
5:04 - Visual Studio Launch Test
5:46 - Creating a New Project in Visual Studio
6:57 - Running the App on Different Machines
7:52 - Sponsor Segment: Quadratic
8:37 - Single-Core Performance Test with Browser
9:00 - Speedometer 3.0 Test
9:53 - Final Thoughts on Performance Across Devices

[elderman]

They should just sell that Dev kit with the finished X Elite chips, that looks sexy

[MrLocsei]

Here's a test idea for evaluating different developer runtimes on identical setups: 

- Create a Windows VM on an X-Elite laptop, and allocate let's say 4 cores to the guest OS.
- Create a MacOS VM on a M-series Macbook, and allocate the same number of cores.
- Run identical tests that use the Python interpreter, Javascript, the Java and .NET (I think .NET framework can target Mac)

Just don't forget to normalise for the CPU frequency. 
This test will reveal the efficiency of these runtimes as implemented for different platforms.

Another setup: instead of a Windows VM on X-elite, create Window VM on another identical Mac, and allocate same number of cores.

Then of course, unleash the SCHWARZENEGGER on them :D

[JohannesLauesen]

again, amazing editing, and amazing speed of informationflow without being confusing... really love your videos.... keep up the style an quality

[MrLocsei]

Virtualisation in case of the Macbook M2 Pro simply means it's running Windows in isolation, no translation or anything else that would hinder performance of the code is happening. So the performance is almost literally what Windows ARM binaries would run on Apple silicon if Apple would've enable Bootcamp. So basically if your host OS is ARM, and the guest OS is also ARM, the "virtualised" binaries should run at native speed. The only difference would be attributed to compiler optimisation - I imagine Microsoft's ARM compiler back-end to target Qualcomm's specific setup (eg. for example make certain assumptions like cache topology or core/cluster allocation strategy of the OS). Certain optimisations could make some difference. Also, Windows' thread dispatched (can't remember it's actual name) / thread scheduler would work differently from MacOS. But by en-large, ARM guest code inside an ARM host OS, like what you're testing, should roughly run at full speed. (with the caveat that MacOS virtualisation framework does not expose the E-cores, at least not on my M1 Pro)

The test is simple: run Geekbench 6 or Cinebench on the Macbook M2, then in a MacOS VM (I tried it with UTM/Qemu) then run Windows 11 ARM on the same M2 macbook with say VMWare Fusion. You should all get roughly the same results.

[antor]

Those video effects of clocks over the laptops are great and provide valuable visual enhancements !

[dgloverau]

Hmmm, you only allocated 4gb to the VM on macOS so suspect that accounted for slower loads and builds

[RaduTek]

I find it hilarious that the Voltera dev kit is basically a Surface Pro X motherboard and a Surface Dock shoved into a box. It does make financial sense for a low volume production.

[MrLocsei]

One interesting thing to think about is why x86 code running with the help of Rosetta, or Prism for that matter, has a performance impact. In theory, it should run at the same speed, as both Rosetta and Prism are fully translating the x86 machine instructions into equivalent ARM before executing even the 1st instruction. In theory, "ideal" x86 code can be transcoded into "ideal" ARM code. The problem is, the compiler already made some decisions and optimisations when compiling say the Swift/C++/Rust code and building for the 86 target - and these are unknown to Rosetta/Prism when they "look" at the block of machine instructions to translate. Moreover, some of these optimisations baked-in by the compiler will have to do with architectural differences such as cache layout and specific memory access optimisations. Still, it is entirely possible, in theory, that the translation to become perfect, and there would be no difference between Rosetta-translated ARM and native ARM. If this doesn't happen, it is because Apple considers Rosetta "good enough" (which is, more than good enough).

[namesurname]

Windows always had problems with the process planner. NPU appeared, but Windows engineers used it to process user data . NPU is best used to synchronize the scheduler of the kernel and scheduler which is built into the processor. When each scheduler works in his own algorithm without coordination, you should not wait for special performance.

[sudeepkansari]

Tell me about machine learning tasks on snapdragon processors and compare it to apple silicon and x86

[Watchandlearn]

The weaker virtualized single core perf of the M2 Max vs the x elite is the reason for the slower build and startup times in vs code for that app since all that is single threaded

[eguif]

Have you installed windows 24h2 in the volterra project?
This version is very focused on our arm

[EduardKorenschi]

Maybe I got it wrong, but on a very short screenshot I've seen, while switching, on the VM on Mac, from 4 to 8 cores, that the memory allocated to the Windows VM was 4 GB. That wouldn't pose a problem for a CPU, but not memory intensive benchmark, like the first one was, but would totally affect the outcome of a VS Code one, with it's memory hog of a Chrome(ium) core. Maybe you could repeat the benchmark with a 8 (10) cores and something like 16GB allowed for the VM. That would make it a bit fairer. I was hoping for the Mac to get it's arse handed over, but it looks different.

[theperfguy]

Problem with MacM(1234) and Windows VM is the price for Parallels. If it was one time payment parallels was probably a better choice for non-gaming, professional Windows (Arm). But that subscription makes it a less attractive option..
Unless you are doing a business (charge it to the business), or your employer is paying for it...

[georgekopadze]

Hey Alex, thanks for great content you always do. Surface you are using in this video, is it 16GB? It’s interesting how 32gb is working

[danhetherington]

Great to see quick builds for development. Other than running the local db from VS what options would you recommended to connect to a sqlserver instance on snapdragon?

[filecore]

Love my Dev Kit 2023. My daily driver and easily good enough for what I do.

[l.bogdan]

Judging by your lastest benchmark videos, Apple now has a serious growing competitor :D. No the time to get lazy on chip advancements :))

[melahi]

Alex, could you please provide the link to the python code of the mandelbrot algorithm are you using for benchmark? Beacuse, of course there are a bunch of mandelbrot implementations in internet. I am pretty interested in the algorithm from the mathematical -numerical point of view.
Thanks in advance.

[namejust]

3:56 i think you set too little ram at virtual machine. it says only have 4gb. even though you can set to 16gb. may i ask why tho?