Skip to content

IN DEVELOPMENT · NO CHIP HAS BEEN TAPED OUT

Why robots do not need the leading edge

The leading edge of chip manufacturing buys density. A robot's own silicon needs certainty: a loop that closes on time, a joint that is driven, a machine that can prove what it is. Mature nodes are enough.

· 4 min read

IN SHORT

A robot's own silicon closes a control loop on time, drives a joint, reads the sensors beside it, and proves what the machine is. Those jobs need certainty, not density, so they do not need the leading edge of chip manufacturing. Mature process nodes are enough, provided the heavy AI work is served from inside the building by a pool such as a Czar Grid.

Most of what is written about chips is about the leading edge: the newest manufacturing processes, with the smallest features and the most transistors in a given area. It is easy to assume that every serious chip belongs there.

BramHA starts from a different view. A robot's own silicon needs certainty, not density, and mature process nodes are enough for it. The program is in development and no chip exists. This article sets out the argument.

What the leading edge buys

Each new generation of chip manufacturing fits more transistors into the same area. More transistors mean more computing in the same space, and usually less power for each unit of work.

That is what large AI models need. Perception, decisions, and adaptation run on models that consume as much dense compute as they are given. For that work the leading edge is not a luxury. It is the point.

The leading edge also has costs. It is expensive to design for. Few manufacturers offer it. And its processes are new, so there is less history behind them.

What a robot's own silicon does

Now consider what the silicon on a machine has to do once the heavy AI work is somewhere else. BramHA is being designed for four jobs.

Control: close a loop on time. Motion: drive a joint or a motor. Sensing: read position, current, and force where they are measured. Identity: prove what the machine is and what it is running.

None of these is short of transistors. A control calculation is small. A drive switches power. A sensor reading is a measurement. A root of trust is a matter of careful design. The page at /chip describes the four jobs.

Certainty, not density

What these jobs are short of, if anything, is certainty.

A control loop needs its answer inside a fixed beat, every cycle, in the worst case and not only on average. A drive has to behave the same way across heat, noise, and years of use. A part built into a machine has to be obtainable for as long as the machine is made and serviced.

Density provides none of that. Certainty comes from how a chip is designed and from how well its manufacturing process is understood. Mature process nodes have long histories. Their behavior is well characterized, and many manufacturers offer them. Determinism over density goes through what a control job asks of a chip.

So the thesis is not that mature nodes are as good as the leading edge. It is that the leading edge is good at something these jobs do not need.

Where the heavy work goes

The argument depends on one condition. The heavy AI work has to be served from somewhere.

In Element 31's family that place is a Czar Grid: an inference pool inside the building that serves every machine in a facility over the site's own network. The pool serves perception, decisions, and adaptation. The machine keeps the work that cannot wait and cannot fail: control, motion, stops, safety.

The Grid is not a controller and does not sit inside any real-time or safety loop. The split is clean in both directions. The dense compute stays off the machine, and the loop stays off the network.

Without a pool, the argument changes. A machine that must carry its own large models needs dense silicon on board, and this program is not designing that.

The cost of the choice

The choice is a trade, and the costs are real. Mature nodes take more area and more power for each unit of compute. A chip built on one has no place for heavy AI work. And a small first design, a real-time controller with a root of trust, does little.

What mature nodes cost you sets those costs out in full.

The program accepts them because of what the jobs are. It also has to be said that the choice is untested by the program itself. No chip has been taped out. The process and the foundry have not been announced, and there are no performance, power, or price figures. What exists today is the reasoning, and it is offered as reasoning.

Questions this raises

Do robots not need powerful AI chips?
The heavy AI work does need dense compute. Element 31's position is that it is served from inside the building by a Czar Grid, so the machine's own silicon is left with jobs that need certainty rather than density.
Which process node will BramHA use?
The process and the foundry have not been announced. The program says only that the silicon is being designed for mature process nodes.
Is a mature node a compromise?
It is a trade. Mature nodes cost more area and power for the same amount of compute and cannot host heavy AI work. For control, motion, sensing, and identity, those costs are acceptable.

KEEP READING