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.
• 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).
• 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.
No comments:
Post a Comment