The architecture

M16: sixteen states, one cell

Most computers built in the last sixty years reduce the world to two states. M16 is our answer to the question of what changes when a single physical unit can hold and read information as a spectrum, not a stack of switches.

But first, let's start with the basics: how a binary chip multiplies numbers and why that led us to pursue multi-state computing in the first place.

Since we are going to be talking about MATRIX multiplication...

Live demo · 3×3 systolic array

Clock 0 / 7Ready

Each clock moves A one cell right and B one cell down. Their product joins the running total stored in that cell. Open the full lab

Inside the hardware

How a chip multiplies grids of numbers

AI is built on matrix multiplication: multiplying numbers in pairs and adding up the results, over and over. A MAC (multiply-accumulate) unit is the tiny calculator that does this. INT8 means each number is stored as 8 on/off switches. Follow one small calculation through six steps.

Our example: multiply [3, -2] by [5, 4], then add
3 × 5 + (-2) × 4 = 15 + (-8) = 7

Representation	Radix	Digits
Binary	2	32
M16	16	8

STEP 01

Turn each number into a small whole number

In INT8 format, a chip can only store whole numbers from -128 to 127. Before any maths happens, every number is converted to fit that range. Our example numbers are already whole and already in range, so they stay the same: 3, -2, 5 and 4.

When it gets harder

Real data is full of decimals, like 0.75. To store it, the chip multiplies everything by a fixed number (say 100) to get 75, and remembers that number so it can undo it later. Anything that does not land neatly gets rounded, which loses a little detail. Numbers bigger than 127 do not fit at all.

STEP 02

Write the numbers as on/off switches

Inside the chip, each number is a row of 8 tiny switches called bits. Every switch is either off (0) or on (1), and the pattern of switches is the number. Negative numbers use a flip-and-add-one trick so the same circuits can handle them.

Number	Switches (bits)
3	`00000011`
-2	`11111110`
5	`00000101`
4	`00000100`

When it gets harder

More precision needs more switches for every number. A real calculation also moves millions of these numbers between memory and the chip, and just fetching them takes time and energy.

STEP 03

Multiply one pair

Multiplying in binary is copy, shift and add. The chip makes one shifted copy of the first number for every switch that is on in the second number. Shifting a number one place to the left doubles it. For 3 × 5, the number 5 has two switches on, worth 4 and 1:

3 × 4 = 12
3 × 1 = 3
12 + 3 = 15

The second pair, -2 × 4, gives -8. An answer can need twice as many switches as the numbers that made it, so products are stored in 16.

When it gets harder

Bigger numbers need more switches, and more switches mean more copies to add. An 8-switch multiplier adds up to 8 copies. A 32-switch multiplier adds up to 32 and needs roughly 16 times the circuitry. Decimals add extra steps to keep the decimal point in the right place.

STEP 04

Add it to a running total

The chip keeps a running total, called the accumulator, and adds each answer to it as it arrives. The total is stored in 32 switches even though the inputs use only 8, because totals grow quickly and need more room.

0 + 15 = 15
15 + (-8) = 7

When it gets harder

Real lists are far longer than two pairs. Adding up 4,096 pairs can reach tens of millions, which is why 32 switches are needed. Every addition also has to pass a carry, like carrying a 1 in school maths, across all of those switches.

STEP 05

Repeat for every row and column

A matrix is a grid of numbers. To multiply two grids, take one row from the first and one column from the second, multiply the pairs and add them up, exactly as in the last two steps. That gives one number in the answer grid. Then move on to the next row and column.

Our example is a single row and a single column, so there is only one answer: 7.

When it gets harder

Real AI grids are huge. Multiplying two grids that are 1,000 by 1,000 gives a million answers, and each one needs 1,000 multiply-and-adds. That is about a billion operations for one multiplication, and an AI model does this again and again. Chips place thousands of MAC units side by side to keep up.

STEP 06

Shrink the answer back to 8 switches

The next stage expects 8-switch numbers again, so the 32-switch total is scaled down and rounded to fit. If it is too big for -128 to 127, it is cut off at the limit. This is called saturating. Our answer, 7, already fits, so it passes through unchanged.

When it gets harder

With decimals, this step also needs a scaling multiply to undo the scale from step 1, plus more rounding. Every rounding and cut-off throws away a little information, and those small losses add up across the many layers of a model.

Why this matters: Small whole numbers make this look easy. Real AI and physics simulations use decimals, large numbers and enormous grids, so every step of the calculation gets bigger, slower and hungrier for energy. Different number formats trade performance against accuracy. Integer formats are faster but less precise. Floating point handles decimals more precisely, but it is slower.

The trade between work and accuracy is the boundary M16 is built to move.

Explore 16 states

Drag the slider from M0 to M15 to move one stit through its sixteen states. For comparison, binary has no single unit that can do this, so it has to line up separate on/off cells: one cell for the value 0, and four cells by the value 15.

MODE 00 / 15

M16 stateM0

Binary, 1 cell0

A stit is not four bits. It is one unit, in one of sixteen states. Four bits are four separate cells, each on or off, that only mean a number when read together. A stit can tell apart as many values as those four cells can, but it does it alone, as a single new kind of unit.

Binary vs. M16

Not a faster binary chip...
a completely new unit of computation

M16 isn't a marketing wrapper around standard logic. It replaces the bit with the stit, which changes what a single unit can represent and what kinds of problems it's naturally good at.

	Binary	M16
States per unit	2 (on / off)	16 discrete states. We have dubbed this a "Stit"
Best suited for	Boolean, Easier component design	Physical simulation, Data dense mathematics

Where this goes

From accelerator to full multi-state compute

The M16 accelerator is the first step, not the destination. Here's how we see the path from here.

Where We Are

M16 accelerator

A dedicated AI inference accelerator built around the 16-state primitives. Our first product, and the proving ground for the architecture.

In development

What Comes Next

General arithmetic multi-state cores

Consumer devices leveraging general arithmetic multi-state cores.

Planned

Where We Are Going

General-purpose computer

The long-term goal: a dynamic multi-state and continuous computing platform.

Vision

M16: sixteen states, one cell

How a chip multiplies grids of numbers

Explore 16 states

Not a faster binary chip...a completely new unit of computation

From accelerator to full multi-state compute

M16 accelerator

General arithmetic multi-state cores

General-purpose computer

Follow the build

Not a faster binary chip...
a completely new unit of computation