Re: Windows-on-ARM Laptop Is A “Frequently-Returned Item” On Amazon

On Sun, 3/23/2025 7:17 AM, Joel wrote:

CrudeSausage <crude@sausa.ge> wrote:

On 2025-03-22 10:08 p.m., Paul wrote:

On Sat, 3/22/2025 5:55 PM, Lawrence D'Oliveiro wrote:

 

It’s clear Windows users have no clue about this Windows-on-ARM thing
that Microsoft keeps trying to push. They just expect their software
to work. But ARM-based Windows machines still require too many
workarounds and suffer too many limitations, and the users are having
great difficulty seeing the point to them.
>
<https://www.tomshardware.com/laptops/snapdragon-x-powered-surface-laptop-7-gets-frequently-returned-item-warning-on-amazon>

>
Microsoft does have a translater, to run Win32 code on ARM.
That's what is on the Snapdragon device.
>
    "What is Prism?
>
     Prism is Microsoft's emulation technology that enables x86/x64
     applications to run on Windows PCs with Arm processors, such
     as Surface Pro 11th Edition, Snapdragon processor;
     Surface Pro 9 with 5G, Surface Pro X, and Surface Laptop 7th Edition,
     Snapdragon processor. It seamlessly translates app code to run on
     ARM architecture, optimizing performance, and reducing CPU usage
     to ensure a smooth user experience on devices powered by
     Snapdragon X series chips."
>
Google is playing up right now, but I gather that isn't working
all that well. Some installers can "detect" they're running on the
wrong platform.
>
One person using one of those products, experienced good performance
at first (right after the OOBE), but as soon as some updates came
in, the emulator performance cratered.
>
Summary: "Safer to test the emulator on a Raspberry PI than spend $3K and return it"

>
Thanks for that, I had no idea that Microsoft actually came up with
something to convert. Of course, it doesn't seem to work all that well
if people are returning their machines the way the press says they are.
It's not like the press would lie or anything, right?
>
I can't help but notice the stellar reviews the supposedly returned
devices are getting.

 
 
It's clear why Microsoft would use x86 emulation with ARM, countless
reasons, but who cares about their Copilot bullshit, put Linux for ARM
on that mother fucker.

Some day, you'll be able to run an AI locally.

The idea is, while you can run multiple video cards,
like the two on the left, you get a better result
if a single card has more RAM on it. In addition to
video cards, some people run "computer clusters" with
crappy network wiring between them. No one mentions
how fast those go. The purpose of me telling you this,
is to discourage buying a shitload of video cards and
hoping it will work well. So far, it would appear we
need to wait for the next generation of video cards
with HBM3 memory on them. I feel better now, that I did not
buy certain video cards L-)

        16GB RAM       16GB RAM                   32GB RAM
        4000 shaders   4000 shaders    versus     4000 shaders
          ^              ^
          |     PCIe     |                        7X faster for AI
          +--------------+

The models are tuned sequentially, and at a certain size.
If a group of experts of size 35GB is on offer, then the
single video card of 48GB size might be able to run it.
The AI is laid out as linear stages. Load and run Step 1.
Load and run Step 2. Load and run Step 3.

     +--------+     +--------+    +--------+
     | Step 1 |     | Step 2 |    | Step 3 | ==>
     +--------+     +--------+    +--------+

Which isn't exactly like how the human brain works.

Even if the hardware had more room, the way it works
might still be discrete steps.

The first step, is strategy planning, for whatever
other steps are required. Loading the video cards would
be from a PCIe Rev5 NVMe. Your question might use the
math module, or the studio art model (for drawing pictures).

But you can test them now, and see if there are good at
anything. At the current time, you do not come away
from the experience, thinking the thing is all that useful.

For example, I asked it to make a list of 2025 internal combustion
cars in the SUV style, and it made a list, but the list
was no better than a regular web site would offer for
such a class of things. If you asked a human
such a question, they would point out which models had
manual heating controls :-) Not so with the AI. Dumb as
a post.

   Paul