Determinism over density
An AI accelerator is asked for as much computing as possible. A control chip is asked for the same result, at the same moment, for years. The two requests lead to different silicon.
· 4 min read
IN SHORT
A control job asks a chip for predictable timing, robust behavior, and a long supply life. An AI accelerator asks for density: as much computing as possible in a given area and power. The two pull apart, and a control chip gains little from the leading edge. That is why BramHA is being designed for mature process nodes. No chip exists.
Two chips can both be called processors and have almost nothing in common in what is asked of them. An AI accelerator and a control chip are a clear case.
BramHA is being designed for the second kind of job. The program is in development and no chip exists. This article compares the two requests and explains why they lead to different silicon.
What an AI accelerator asks
An AI accelerator runs large models. Its work is a very large amount of arithmetic, and the measure of success is throughput: how much of that arithmetic it completes for the area and power it uses.
So an accelerator asks for density. More transistors in the same area mean more computing. This is what the leading edge of chip manufacturing provides, and it is why accelerators are built there.
An accelerator can tolerate some variation in timing. If one answer takes a little longer than the last, the result is still correct. What matters is the total.
What a control job asks
A control job asks for three different things.
Predictable timing. A control loop measures, calculates, and commands on a fixed beat. The answer has to arrive inside the beat every cycle. The measure is the worst case, not the average. A chip that is usually fast and occasionally late is a poor controller.
Robustness. A control chip sits inside a machine, near motors that switch power, in heat, vibration, and electrical noise. It has to behave the same way through all of it, for the working life of the machine.
Long availability. A robot maker designs a part into a machine and then builds and services that machine for years. The part has to remain obtainable over that time. A chip that is replaced by a new generation every short while is a cost to everyone who built it in.
None of the three is a request for more transistors.
Why the two pull apart
Density and determinism are not opposites, but the pursuit of one tends to work against the other.
The techniques that raise throughput make a processor's timing depend on what it happened to do just before. The average improves. The worst case becomes harder to bound. For an accelerator that is a good trade. For a controller it is the wrong one.
A control design goes the other way. It prefers a simpler processor whose slowest cycle is known, and it gives up average speed to get it. Predictable timing comes from how a chip is designed, not from how small its transistors are.
Robustness and availability point the same way. Mature process nodes have long production histories, so their behavior is well understood. Many manufacturers offer them, and they tend to stay in production. What a mature process node is explains the term.
So a control chip gains little from the leading edge and gives up a good deal to be there: cost, a narrow choice of manufacturers, and a process with less history behind it.
What this means for the first design
The first E31 design is a real-time controller with a root of trust. It is deliberately small. It is not a vision processor and it is not an AI accelerator.
That follows from the comparison. The controller is being designed so that the loop closes on the chip, which is a determinism job. A control loop that closes on the machine describes it.
The dense work has a place of its own. Perception, decisions, and adaptation are served from inside the building by a Czar Grid, which is not a controller and does not sit inside any real-time or safety loop. Each side gets the hardware its job asks for.
The honest limits
Choosing determinism over density is a trade. Mature nodes take more area and more power for each unit of compute, and a chip built on one has no place for heavy AI work. A small first design does little.
It also has to be said that determinism is a design goal here, not a result. No chip has been taped out. The specification of the first design is not published, and there are no timing, power, or price figures. The process and the foundry have not been announced.
Nor does predictable timing make a part a safety system. The chip may run a vendor's control loop. The safety case is the vendor's, and no safety or functional-safety certification is claimed.
Questions this raises
- What does determinism mean for a chip?
- It means the chip does the same work in the same bounded time, every time. A control loop depends on that. A result that is usually fast but sometimes late is not good enough.
- Could a fast AI chip also run the control loop?
- A dense chip is built for throughput, and the features that raise throughput tend to make timing harder to bound. Control is better served by a simpler design whose worst case is known.
- Does BramHA publish timing figures?
- No. There are no performance figures. No chip has been taped out, and the specification of the first design is not published.
KEEP READING