Chip, litografia, efficienza, quantistica: più calcolo con meno watt, e cosa ne faremo.
Il calcolo è la risorsa nascosta di tutto il resto: senza chip efficienti niente intelligenza artificiale, niente veicoli autonomi, niente reti intelligenti. La rubrica segue la litografia che scende — 2 nanometri e oltre —, le architetture che fanno più operazioni per watt, le memorie, le interconnessioni e il calcolo quantistico quando produce risultati verificabili invece di promesse.
Leggiamo le roadmap di TSMC, Samsung, Intel e delle fonderie cinesi, le pubblicazioni di architettura, i benchmark indipendenti. L’angolo è costante: l’efficienza — il calcolo per unità di energia — conta più della potenza bruta, perché è lei a decidere che cosa funzionerà davvero, dal data center al telefono. Ogni articolo dà i suoi ordini di grandezza e cita le sue fonti.
ALICE observed suppressed neutral-pion production in oxygen–oxygen collisions, matching a signature seen in lead–lead collisions.
The MV3564 and MV3514 combine a VCO and multi-PLL synchronization on TSMC’s 40nm RF CMOS process.
The chip supports optical-switch products with up to 400 by 400 ports.
A calcium-40 ion served as a microscopic sensor inside an electromagnetic trap.
UT Austin and TSMC tested magnetic-memory chips for AI tasks, including inference and neural-network training.
Researchers in Shanghai used Step-Eva to grow single-crystal metal contacts directly on 2D semiconductors.
KIMS used amorphous silicon, tellurium and an electric-field layer on flexible polyimide.
Waseda researchers removed key constraints from a slime-mold model of the traveling-salesman problem.
CUDA Python 1.0 bundles several packages under a shared versioning strategy.
The new GPU scheduler allocates compute time based on usage instead of strictly by queue order.
The material is only a few molecules thick and changes its electronic properties with light.
Shunkai launched at the Institute for Molecular Science with about 50 qubits and a QPU from Infleqtion.
The AGI CPU has up to 136 Neoverse V3 cores across two dies.
Wits and Bordeaux researchers encoded data in light topology, a property that can survive severe beam distortion.
Tests on Semianalysis’s InferenceX benchmark put Jalapeño ahead of Nvidia Blackwell on two efficiency measures.
The M6 is Apple’s first 2nm chip, with a 12-core CPU, 12-core GPU and up to 32GB of unified memory.
The method screens color centers without the intensive simulations normally used to assess their optical losses.
The chip creates a structured microcomb: many precisely organized light frequencies at once.