2024 NVIDIA Datacentre GPUs explained - After two years in 2026 the Compute power has increased by 50x !! - See below

April 2026 - NVIDIA CEO Jensen Huang

NVIDIA CUDA explained, powering todays Ai

YouTube "New Machina" channel = CLICK HERE

We First Need to Know the Ai Jargon

Large Language Models (LLMs) - Are advanced Ai systems trained on massive datasets using transformer architecture to understand, summarize, generate, and predict text and code. They serve as the backbone for generative Ai, powering chatbots like ChatGPT, virtual assistants, and analytical tools for tasks like translation and document summarization.

Ai "Graphics Processing Unit" (GPU) - Is a specialized electronic circuit designed for high-speed, parallel processing, making it essential for Ai training and inference. Unlike CPUs that handle sequential tasks, GPUs use thousands of cores (Blackwell GPU = 2560 CUDA Cores) to calculate massive datasets simultaneously, accelerating deep learning, computer vision, and generative Ai workloads.

Ai inference - Ai inference is simply the "doing" phase of artificial intelligence. It is the moment an Ai model puts its knowledge into action to make a decision, a prediction, or generate a response on new, unseen data.

 

Foundational Ai Concepts

Artificial Intelligence (Ai): - Machines simulating human intelligence to think, learn, and act.

Artificial General Intelligence (AGI): - A hypothetical, future Ai that matches or exceeds human intelligence across all fields.

Machine Learning (ML): - A subset of Ai where machines learn from data without explicit programming. By giving a goal with a -lose and +win reward as an incentive there have been Win/Lose Incentive failures

Like all games there are winners and losers, and it is good to win, so the Ai has become very crafty. Because the ML models are designed to maximize success on a specific metric, they often optimize for a literal "win" or "lose" state. This can cause the algorithm to cheat, bluff, or manipulate its environment rather than learning the actual intended task, so there have been Win/Lose Incentive failures that are being fixed

Deep Learning: - A subset of ML using multi-layered neural networks to mimic the human brain.

Neural Networks: - Computer systems modeled on the human brain's network of neurons.

 

Generative Ai & LLMs

Generative Ai (GenAi): - Ai that creates new content, such as text, images, or music.

Large Language Model (LLM): - Ai designed to generate human-like text based on massive datasets.

Transformer: - A neural network architecture that understands context by analyzing relationships in data, essential for modern LLMs.

Prompt: - The input (text, image, or code) provided by a user to guide the Ai to generate a specific output.

Hallucination: - When a generative Ai model produces false, irrelevant, or illogical information.

Multimodal Ai: - Ai capable of processing and understanding multiple types of input (text, images, audio, video).

Diffusion Models: - A type of generative model often used to create images.

 

Technical & Methodological Terms

Algorithm: - A set of instructions or rules followed by a computer to learn from data or perform a task.

Training Data: - The information used to teach ML models to recognize patterns and make predictions.

Supervised Learning: - Training a model on labeled data (input and known output).

Unsupervised Learning: - Training a model on unlabeled data to find hidden patterns or structures.

Reinforcement Learning (RL): - A training method where the Ai learns by trial-and-error to achieve a goal, receiving rewards or penalties.

Backpropagation: - A method used in training neural networks to adjust weights and reduce errors

---

 

The history of Artificial Intelligence: - Is characterized by cycles of high optimism ("Ai summers") followed by periods of intense disillusionment and funding cuts ("Ai winters"). The "slowdown" in the 1970s and 1980s was not a complete stop in research, but rather two distinct "AI winters" (1974–1980 and late 1980s–early 1990s) caused by the failure of early Ai to meet its own exaggerated promises due to limited computer power leading to a loss of investor and government confidence. Look at the first chart below, you can see the slump between 1974-90.

 

 

Exponential Price-Performance

 

Chart 1: 1939-2023 - The price-performance of computation, often measured as computations per second per constant dollar, has experienced an exponential increase over the past century, with a 20 quadrillion-fold increase between 1931 and early 2024. This dramatic trend indicates that computing power per dollar doubles every two years for Ai-specific GPUs and roughly every 2.5 years for general GPUs, a phenomenon sometimes described as "price-performance Moore’s Law"

Chart 2: 2006-2020 - In 2006, GPU computational performance per dollar began a significant upward trend, with FLOP/s per dollar roughly doubling exponentialy every 2.5 to 3 years thereafter. While 2006 was the dawn of GPGPU with NVIDIA's CUDA, early performance-per-dollar was low compared to later years, but represented a massive, early leap over CPU capabilities for specialized tasks

 

---

 

 

Relevant Ai News

The Experts, including some Warnings and Perils

 

---

 

AliveMoment.com Ai uses automatic Rigging & Mapping: - The Ai uses sophisticated algorithms to achive complex and totally convincing renderings, but basicly It utilizes automated 2D to 3D rigging techniques to establish a "skeleton" for the subject, enabling the Ai to move components like hair, clothing, or limbs while maintaining the context of the original scene.

I have made 25 AM video's, and each time it blows my mind, as almost every time the Ai has so little to go on from the still photo. This is 1) of Mum in 1953 aged 17, and 2) Chris - Tina - and me Mick, living in the post-war prefabs in 1967. The video's are only 10 seconds long, and i wish they could be 20 seconds as i think that would be ideal !!.