The Bible of AI

When does cooling an AI processor from inside the silicon improve efficiency?

When does cooling an AI processor from inside the silicon improve efficiency?

Engineers can now etch cooling channels into the back of an AI processor and circulate liquid close to its hottest regions. Tests on a working CPU and an NVIDIA GH200 show what this changes for sustained clock speed, GPU temperature and nearby high-bandwidth memory. The electricity question requires a wider view. A retrofit at Lawrence Livermore reveals how much depends on which fans slow down and where the heat finally goes. A separate Berkeley Lab study shows how changing the cooling plant alters the projected savings. Together these results explain when a data centre might serve more AI requests within an existing power limit, which parts of a liquid-cooled rack it could reuse, and what manufacturers must establish about sealed channels, filtration and reliability before deploying the technology across a fleet.

When does cooling an AI processor from inside the silicon improve efficiency? Leer más »

Can gallium arsenide replace silicon? From a working processor to the optical links that carry AI data

Can gallium arsenide replace silicon? From a working processor to the optical links that carry AI data

Engineers built a 32-bit experimental processor in gallium arsenide. Today a more familiar place to find the material is inside the optical transmitter that carries data between computing systems. What determines which role it can take in AI? This review follows electrons from a crystal into a transistor and then into a circuit with registers, wiring and a clock. Studies from Japan, China, Korea and the United States show what changes when gallium arsenide forms the channel, supports another semiconductor or shares a chip with silicon logic. Optical experiments in Sweden and China reveal how the distance between machines and the choice of signalling affect the data rate. A French–German neural experiment gives the material a third role. Together these results identify the measurements needed to establish an advantage in a complete AI workload.

Can gallium arsenide replace silicon? From a working processor to the optical links that carry AI data Leer más »

What can silicon carbide do for AI? From efficient power to extreme-environment computing

What can silicon carbide do for AI? From efficient power to extreme-environment computing

What can silicon carbide contribute to artificial intelligence: a more efficient power supply, computation beside a hot sensor, or an artificial neuron? This article follows those distinct routes through original research and industrial documents. It explains why equal numbers of silicon and carbon atoms do not mean equal masses, and why shorter atomic bonds do not guarantee a faster processor. Fabricated logic gates and memory cells establish building blocks; NASA’s 2026 processor paper distinguishes a SiC design from a robot demonstrated with a silicon FPGA at room temperature. Cryogenic neuron experiments open another specialised direction, with refrigeration still part of the system’s energy account. The practical question is which function SiC can perform, under what conditions, and with what measured benefit to a complete AI workload. Power conversion offers a clear role; replacing today’s AI processors requires much more evidence.

What can silicon carbide do for AI? From efficient power to extreme-environment computing Leer más »

R0:e81127dd3f66450ac34b82915cf63584-An ECG biomarker for sudden cardiac death discovered with deep learning

An ECG biomarker for sudden cardiac death discovered with deep learning

A routine electrocardiogram can reveal a disturbance in rhythm while it is happening. Could the shape of an apparently ordinary beat also help identify someone at risk of sudden cardiac death a year later? Ziad Obermeyer and colleagues linked Swedish ECGs to death certificates, then examined what their model had learned. When they altered a beat to raise its predicted risk, the descending edge of one wave became less abrupt in lead aVL. That observation gave them a feature they could measure in actual recordings. The Swedish results show both the promise and the limits of the approach. A small group had a markedly elevated event rate, yet many ECGs preceding a death fell below the chosen threshold. Records from the United States and Taiwan add evidence about ventricular arrhythmias and cardiac arrest. Whether offering defibrillators to people identified through this signal would prolong their lives remains an open clinical question.

An ECG biomarker for sudden cardiac death discovered with deep learning Leer más »

Dory1.0-4B and the value of a second pass

Dory1.0-4B and the value of a second pass

Can a language model solve a harder problem by applying the same learned transformation a second time? NVIDIA’s Dory1.0-4B makes this question experimentally tractable. Its central block can process the same sequence at different depths while the model’s four billion parameters remain fixed. NVIDIA reports a substantial gain from the second pass on software repair and reasoning over long documents. Further passes yield smaller gains on average, and factual accuracy changes little on one knowledge test. This assessment examines the published weights, training design and evaluation results alongside measurements from people running the model on their own hardware. Experiments with llama.cpp and a Rust inference engine illustrate how precision, memory management and the choice of execution depth affect speed. These comparisons also show why the number of generated tokens, the time required to answer and the quality of the answer deserve separate measurements.

Dory1.0-4B and the value of a second pass Leer más »