The Bible of AI

Carver Mead and the Eye Made of Silicon

Carver Mead and the Eye Made of Silicon

When Carver Mead received a lifetime award for neuromorphic engineering, he used his speech to credit the student who had persuaded him to build a silicon retina. Misha Mahowald’s question suited an engineer who had grown up beside a hydroelectric plant and later helped students design chips of their own. Could the connections between electronic components do some of the work of seeing? Their experimental retina compared light across neighbouring points. A second collaboration, with Richard Lyon, produced a circuit that separated sounds according to frequency. This biography follows Mead from the design rules he developed with Lynn Conway to experiments that a reader can picture with a pencil line and a sheet of paper. It also follows those questions into a listening device and a table-tennis robot, where electrical power and the time taken to find a moving ball can both be measured.

Carver Mead and the Eye Made of Silicon Leer más »

What does an AI chip do between events? Inside neuromorphic vision

What does an AI chip do between events? Inside neuromorphic vision

An event camera watching a doorway can report a change at the edge of a moving coat while the stationary doorframe generates little traffic. What does a neuromorphic chip do with that difference, and how much electricity can it save over a quiet hour? Two experiments offer a way to answer. Speck puts a vision sensor and a spiking processor on one chip. Its gesture recognition measurements show how filtering events changes both power and accuracy. A separate object detection study compares Intel’s Loihi 2 with a Jetson Orin Nano and finds different winners for energy per inference and detection rate. Following a single crossing from a pixel to an alarm reveals what those comparisons leave open. A practical test would count timely detections and false alarms while measuring energy at the supply for the complete device.

What does an AI chip do between events? Inside neuromorphic vision Leer más »

What happens when Canon’s wafer stage brakes? Recovering energy to make memory for AI

What happens when Canon’s wafer stage brakes? Recovering energy to make memory for AI

A lithography scanner spends electricity moving a silicon wafer from one exposure field to the next. What happens to that energy when the stage slows down? Canon’s FPA-6400ES6a captures part of it for reuse. Canon reports lower electricity consumption per wafer than its preceding KrF scanner and specifies 400 wafer passes per hour under a defined exposure recipe. We follow the energy through one braking cycle and examine why a faster machine can use less electricity per wafer yet draw more power over a busy shift. The distinction matters for factories making memory for AI accelerators. A useful comparison would measure electrical energy at the machine inlet, alignment on the patterned wafers and the number of memory dies that pass electrical tests. Those measurements would show how much of the benefit reaches the supply of usable memory.

What happens when Canon’s wafer stage brakes? Recovering energy to make memory for AI Leer más »

Where does power enter a nanosheet transistor? The case of TSMC A16

Where does power enter a nanosheet transistor? The case of TSMC A16

Where does the supply current enter a transistor built from stacked sheets of silicon? TSMC’s A16 process brings power to the device from the back of the wafer. That change can leave more room above it for the wires that carry signals between logic cells. It also explains how a manufacturer can report higher chip density while preserving gate density. This review follows the contact from the supply rail to the transistor and checks the comparison with TSMC’s N2P process. Two reports give the same range for a speed gain while holding different electrical quantities constant. A separate simulation from imec shows why the choice of contact can matter more as cells become shorter. For an AI accelerator, the decisive questions are how much of a workload waits for logic, how much waits for memory, and how much energy the whole system spends on each completed inference.

Where does power enter a nanosheet transistor? The case of TSMC A16 Leer más »

Jack Kilby and the Circuit That Fit on a Sliver of Germanium

Jack Kilby and the Circuit That Fit on a Sliver of Germanium

When Jack Kilby switched on an oscillator at Texas Instruments in September 1958, a sliver of germanium produced a steady electrical signal. Its fine gold wires tell another part of the story. They made the demonstration work, but each connection still had to be made by hand. Kilby’s route from a Kansas ice storm to that Dallas laboratory helps explain why he thought an entire circuit could be built from semiconductor material. Robert Noyce, Jean Hoerni and a team at Fairchild found a way to put the wiring into the manufacturing process itself. This biography follows those distinct achievements into a calculator that printed its answers on paper and, decades later, into the processors that run neural networks. Across those devices a practical question keeps returning. How far must an electrical signal travel to do useful work?

Jack Kilby and the Circuit That Fit on a Sliver of Germanium Leer más »

RE LightGBM M3 and the retrospective audit of Italian public contracts

RE LightGBM M3 and the retrospective audit of Italian public contracts

Which completed public contracts deserve a closer look when auditors can review only a small share of the files? RE LightGBM M3 addresses that question using Italian procurement records, administrative court decisions and selected grounds for contract closure. The research team has released the fitted decision trees, its rules for assigning labels and the scores from held-out contracts. In the published comparison, the highest-scoring one per cent contained roughly one fifth of the known positive cases. Yet the same model showed little ability to distinguish positive from negative outcomes across the labelled cases as a whole. This profile examines how both findings can hold, what a successful administrative appeal establishes, and what an audit authority would need to measure before using the ranking to allocate staff time. Its relevance lies in a question that the public evidence makes it possible to test.

RE LightGBM M3 and the retrospective audit of Italian public contracts Leer más »

Goose versus Rust and C++: the case for keeping data in place

Goose versus Rust and C++: the case for keeping data in place

Goose lets a growing array keep its address and a whole group of objects disappear when their work is done. Wouter van Oortmerssen’s language turns those ideas into measurable savings on some workloads. The gap narrows when Rust and C++ use equally careful layouts. We follow the designer from Amiga compilers to FlatBuffers, examine how Claude Fable helped build Goose, and read the figures behind its strongest claims.

Goose versus Rust and C++: the case for keeping data in place Leer más »

Diamond beyond silicon: what Ōkuma and boron compounds mean for AI chips

Diamond beyond silicon: what Ōkuma and boron compounds mean for AI chips

A new factory in Ōkuma, Japan, brings diamond semiconductor devices closer to manufacture. But the clearest link between diamond and artificial intelligence today starts with a different task: carrying heat away from a silicon processor. This review follows both paths, from the Japanese plant and laboratory diamond transistors to cooling experiments in working devices. It distinguishes boron added to diamond from cubic boron arsenide, boron carbide and a one-layer boron carbon nitride transistor channel. One experiment shows why the quality of the bond can matter more than the thermal conductivity of the crystal. Another reveals how heat and electrical charge can travel through the same specimen without telling us how fast a processor will switch. An AI server already advertised with diamond cooling provides a practical test case: what would an independent laboratory have to measure before claiming lower energy per token?

Diamond beyond silicon: what Ōkuma and boron compounds mean for AI chips Leer más »