Friday, November 20, 2026

CPU, GPU and AI

CPU
Sequential reading of commands

GPU
Parallel reading and writing of DATA parameters as Pixels on the video monitor.

AI Model
Runs not on a single computer but on a Network of Computers in the cloud. 

There is no Computer Language Model but a Pre-trained Learning Model that transfer data through a method called Transformer.

Transformer could read 
CODE with Syntax 
More importantly can read Human Language in blocks of written language.
But from GPT Model 4 in can read Sensory input from a touch screen or a video monitor.
 
Transformer Documents was developed by Google Team of Research  but these guys who left Google and Developed their own AI companies.

Ordinary commuter with 8 to 16 RAM and a moderate GPU.
The graphic card can play games and run video or audio applications.

Neural Network Concept burrowed from how network of neurons in our brain works managing spoken and written language and decision making by humans in a simple and intelligent manner.
It is the base of the AI model.
 
Humans with  Intelligence and Machine with DATA (information)
 
While brain works on Memory, intelligence, emotions and thought processes, the AI Model works on DATA sets available (written language or sensory inputs) as inputs.
No thinking
No emotions
 
But rapid response to unlimited amount of information the GPT 4 Model receives.

This is where the GPU come into operations running parallel with CPU that processes and runs commands in threads (thread is a continuous stream of instructions that CPU handles).
 
The number CPU cores determines how fast the data fed in by GPU is handled sequentially by the CPU. The sequential processing causes bottlenecks.

All these are processed by the RAM which the computer invariably should have and is designated as a volatile memory. 
In Linux terminology the memory blocks were called Pages and in good old days with limited RAM, say 500MB, Linux needed to have SWAP place (partition) to hold memory pages not running currently but may be needed to run a program execution in future tense.

We use 300 millions of visual data set with a good graphic card.

However, DATA can be in millions, billions or in trillions.
 
AI Model needs millions of DATA input where RAM should exceed 64 or more.
That is why some Mini computers working as routers have 128GB RAM. 
I think to play games 64 GB is enough anything more is a waste of money.
 
 LLM
 An LLM (Large Language Model) is a specific type of artificial intelligence (AI) program designed to understand, process, and generate human-like text.
 
Training Data: 
LLMs are trained on massive datasets scraped from books, articles, websites, and code
 Transformer Architecture: 
They rely on deep learning neural networks (specifically transformers) that analyze how words and sentences connect across an entire text
 
 Next-Word Prediction: 
When you give the AI a prompt, it calculates probabilities to guess and generate the most logical next word over and over again
 
These LMMs need powerful graphic cards, namely NVIDIA.
In good old days we never used to call the memory in GB.
We used to call the model of the unit or next generation graphic cards which handled video games better than the previous generation.  
 Now this NVIDIA tensor core graphic cards have series of chips embedded in the unit and they are bundled as number of core
 
A 4×4×4 and 8×8×4 array are multidimensional (3D) data structures or physical grids used in programming, math, hardware design, or visualization.
 
 • 4×4×4 Array (3D Grid / Cube):
    • Represents a 3-dimensional block with dimensions 4 × 4 × 4.
    • Contains a total of 64 elements (4 × 4 × 4 = 64).
    • Common examples include a 4×4×4 LED cube, a 3D voxel grid, or a 3D array in programming languages like C++ (int arr[4][4][4]) or MATLAB.
 
8×8×4 Array (3D Grid / Layered Matrix):• Represents a 3-dimensional block with dimensions 8 rows × 8 columns × 4 layers (or depth of 4).
• Contains a total of 256 elements (8 × 8 × 4 = 256).
• Often used in graphics,
neural network tensor blocks, hardware design, or multi-layered pixel/LED matrices (such as 4 layers of 8×8 grids).
 
 
 
Neural Network needs 650 parameters for Large Language Models necessary for AI. 
At least 10,000 GPUs are necessary for such parallel processing.
AI and GPUs generate enormous amount of Energy and a cooling system is necessary to run them efficiently..

Video Games consume lot of energy depending on the pixel density used.

Fighting China on the Technology Front is Futile

China is fast to adapt in technology front.
BeeLine Mini PC is an example.
Huawei Trade mark and company will soon overtake Intel based Lenova.
China had a Linux operation called Dream OS but it failed but Huawei in less than 2 years has weathered the storm. 
I am not going to adopt Chinese products since I do not use a laptop but my NUC for my work. ACER laptop I bought in 2009 is gathering dust.
Besides, I use Linux and have not used Windows for over 25 years without loss of productivity due to Linux Operating System, Debian.

ToOpy

ToOpy

ToOpPy Linux is a French distribution based on Puppy Linux. The project provides a lightweight distribution which includes many small utilities and can be run either from a live disc or installed on the hard drive.

If I find a guy with a 32 bit computer not working. I would boot one of these images and analyse the problem first hand, with the hardware including Graphic Card and RAM. If I cannot boot Puppy Linux, I am certain the RAM is at fault and I discard the hardware, since finding a RAM card for these obsolete hardware is not worth the trouble.

 

Puppy Linux

 

Puppy Linux

I found “Puppy Collection” in an archive and downloaded a few of my favourites in the past. Most of them were only 32 bit compatible versions.

Puppy Linux is a family of Light Weight Linux distributions that focus on ease of use and minimal memory footprint. The entire system can be run from RAM with current versions generally taking up to about 600MB (64 bit), 300MB (32 bit), allowing the boot medium to be removed after the operating system is running in full swing.Applications such as Abiword, Gnumeric and MPlayer are included, along with a choice of lightweight web browsers and it’s own Puppy Utility for downloading packages.

The distribution was originally developed by Barry Kauler and other members of the community, until Kauler retired in 2013.

The tool Woof can build a Puppy Linux distribution from the binary packages of other Linux distributions. It went up from 0.1 to 1 to 2 to 3 to 4 without a Family name until Wolf Pup 4.3. From then onward, it diverged and proliferated and until 64 bit version came into existence, I tried all of them which I could download as Live CDs.

USB sticks were rare to find, then.

 

32 bit Versions

Wolf Pup

Wolf 4.3 in 2009

Quirky end in 7.0

Racy in 2013, end with version 5.5

Wary in 2013, ends with version 5.5

Precise in 2013, ends with version 5.7.1

Lupu in 2013 and ends with Woof-CE in 2014

Slako in 2013, all these were 32 bit versions.

 

64 Bit Versions

Quirky 7.3 in 2015 ends up with Xerus

Tahrpup 6.05 in 2015

Slacko 2015

Currently, only 5 of them are 64 bit compatible.

FatPup-900 on

Bookworm Pup 64.10.0.2

F96-CE-4

VanillaPup10.6.0

Void Pup.64.22.0

Puppy Releases

Puppy 0.1 18 June, 2003

Puppy 1.0 29 March, 2005

Puppy 2.0 1 June, 2006

Puppy 3.0 2 October, 2007

Puppy 4.3.0 5 May. 2008

Puppy 5.0.0 15 May, 2010

Puppy 5.2.8 4 April, 2012

Puppy 5.3.024 October, 2011

Puppy 5.7.08  March, 2014

Puppy 6.0.5 26 October, 2014

Puppy 6.3.2 21 June, 2016

Puppy 7.5 4 December, 2017

Puppy 8.0 24 March, 2019

Puppy 8.2.1 1 July, 2020

Puppy 9.5 21 September, 2020

TrixiePup64

 

TrixiePup64

TrixiePup64 (11.2/2601) is a lightweight, Debian-based Puppy Linux distribution derived from Debian 13 "Trixie," designed for speed on older hardware. It features 64-bit and 32-bit versions, supporting both Wayland (using Labwc) and Retro X11 environments, while booting entirely into RAM.            It includes APT/Synaptic package management.

Built upon Debian Trixie (Debian 13).

Wayland Version:

Uses the Labwc compositor, SFWbar and zzzFM, featuring Xwayland for compatibility with X11 applications.

Retro Version:

Features a traditional X11 desktop for older systems.

Performance:

Designed to run efficiently on low specification hardware, often running completely in memory for high speed.

Package Management:

Full access to APT and Synaptic Package Manager for installing software.

TrixiePup64 2601:

Released in early 2026, utilising the Linux 6.12 kernel.

TrixiePup64 11.2:

A notable release in late 2025 focusing on updating components from the Trixie source. A variant offering, with specific builds utilising LXQt 2.1, KDE Plasma or Cinnamon, providing a more modern desktop experience while maintaining compatibility with Debian 13.

Root Usage:

Similar to other Puppy versions, users generally operate as root by default. However, some versions like PuppEX allow for a standard user login.

Installation:

Uses "frugal" installs to a hard drive and supports UEFI/non-UEFI systems. This is where            I disagree. It mounted all 41 of my partitions but both frugal installer and puppy installer failed write a copy of sfs file. I was a regular user of Puppy Linux and I knew how Puppy operates.  I even tried on a FAT32 partition. I think it needs a clean hard drive with only Microsoft Windows. It ignores Linux distributions just like CachyOS and Manjaro. I have only 12 distributions in my 500GB hard disk.

I ultimately gave up.

Save Session:

Allows saving changes at shutdown to a folder on a hard drive or USB partition.

Some users consider it less "traditional" than older Puppies (like BionicPup), noting that the Wayland versions can have issues with drive icons and taskbar customisation.

Puppy Easy OS-7.1

Puppy Easy OS-7.1

I have had installed Puppy Easy OS on a USB stick. The latest version 7.5.1 does not boot ob my NUC.

It has lot of problems, basically due to its container technology. Containers can be used for many uses but mainly for Virtualisation.

What it does is to mount something foreign on top of a very basic operating system. What it tries to do is to work without link libraries.

Linux is not resource hungry.

It allocates and prioritise resources as and when needed. All these techniques are mitigated or vandalised by containers.

In good old days strength and beauty of Puppy were it's small footprint of about 100MB. Now, it has expanded to 1000MB and more and is showings it's vulnerability. Mind you, sfs is a long single text file.

It looks like its seems are bursting.

Gnome has a better Virtualisation package called Box Utility. The beauty of Box Utility is it gives one the ability to try individual packages within the Container. It consumes lot of RAM and it need fare amount of hardware capacity as reserved space to run Box.

Flathub is inherently bulky.

Yes, Puppy should move forward with new technology but with extreme caution.

It's package installer is FlatPak and it is painfully slow. I wanted to Install Audacity, KDN Live and my favourite browser Firefox but it is getting frozen in mid air like a jet plane without jet fuel.             I tried Brave browser and the same outcome. I generally have at least 3 browsers working for me in addition to Firefox. They are Dillo and Falkon and Konquorer.

Yes FlatPak is a pain compared to Synaptic Package Manager of Gnome (Debian). I have 8GB of RAM and that is enough for my daily work.

If Easy Pup cannot allocate my RAM in an equitable way, I am not going to recommend it to any guy or girl. Currently, there are no other Puppy derivatives available but luckily I have a collection of old Puppy Images with me, including Ubuntu and Debian derivatives.

Pluses

On the plus side it configures WiFi effortlessly.

Chrome browser is OK.

It allocates entire 8GB USB to a EXT4 File.

Puppy on a USB is my favourite when I am on the move. Thankfully, today is a holiday and I can drop to sleep while these actions are running in the background.

I re-did everything.

It is all due to container philosophy which is pretty slow.

One should install one application at a time.

Flathub does not know how to queue one application after another.

It tries to do everything in one go, hence it gets chocked.

Linux philosophy is one job at a time with less time allocation for less demanding task.

 

BIOS Master Password Generator

 

BIOS Master Password Generator

The biosbug.com and the bios-pw.org are the organizations that give free assistance to unlock the BIOS lock. When one hits the F2 on booting of the locked laptop, it gives a  8 digit number and when that number is entered into a dialogue box provided at their web sites, the unlocking number is generated which one has to enter to unlock the BIOS. Thereafter, one reaches the BIOS Setup and one may leave without a password or add one’s own password.

Since, I forget the passwords, I do not have one. The end result is that I have seven Linux distributions on my old laptop with 320MB hard disk.

This vendor-locked laptop is probably, the very reason I migrated to Linux for good.

I did not have the  proper sense to test it before purchasing and beware if one purchases a  "marked down" gadget even in Singapore.

Nowadays, I use Puppy Linux to Test Drive but before that it was Knoppix. After a while, Windows  Installation was erased with True OS which is defunct now. The True OS Boot CD automatically connects to the Internet, thereby bypassing the USB boot menu.
 If the laptop lacks the CD/DVD ROM, one is totally handicapped. 
I used to have over 100 Linux Boot DVDs. I had to use 6 or 7 CDs to boot Debian 6.      
High capacity re-writable DVDs are not available in Australia.