Why axes matter
A periodic table isn't a list. It's a coordinate system. Every element has a position, and the position tells you something about the element. Atomic number increases going right. Electron shells fill going down. Chemical behavior repeats in columns. The structure carries the insight.
When Mendeleev arranged the elements, he noticed that two properties were enough to organize everything. Atomic mass ordered them linearly. Chemical family grouped them into columns. Two axes. Done.
Computation is richer than chemistry. A single operation has more independent properties that matter. Six of them, as it turns out.
Axis A — Family
The algebraic structure the operation belongs to. What kind of math is it? There are thirty-three distinct families in the stack, from foundational arithmetic to finite fields to semirings to neuromorphic spiking to noncommutative geometry. An operation's family tells you what kind of hardware would execute it most efficiently.
Axis B — Abstraction level
Where does the operation live? The stack has five levels:
- B₀ — pure math (it's an identity or definition)
- B₁ — gate level (implemented in transistors)
- B₂ — microarchitecture (implemented as a pipeline)
- B₃ — instruction set (exposed via an opcode)
- B₄ — software (implemented by code)
An operation's abstraction range tells you where you could build it. Some operations only make sense at one level. Others can live at any level depending on your tradeoffs.
Axis C — Thermodynamic cost
What does the operation have to cost in energy? This is physics, not preference.
- L₀ — reversible, cost floor is zero
- L₁ — measurement, requires entropy source
- L₂ — Landauer erasure, pays
kT ln 2per destroyed bit - L₂+L₃ — erasure plus memory cost
- L₂-max — maximum erasure, the operation destroys almost everything
Axis D — Concurrency topology
How does the operation parallelize? Scalar, vector, matrix, graph, embarrassingly parallel, or sequential. Tells you what shape of hardware fits it best. A matrix operation wants a systolic array. A graph operation wants a mesh. A sequential operation wants deep pipelining. These aren't aesthetic choices — they're dictated by the operation.
Axis E — State topology
What does the operation remember? Stateless, finite-state, accumulating, or feedback. This axis determines memory architecture. Stateless operations need no state. Accumulating ones need a running value. Feedback ones need control loops. Three-quarters of the primitives in the stack are stateless — and they're the easiest ones to build hardware for.
Axis F — Information topology
How does information flow through the operation?
- Bijective (1-to-1) — no information lost. 21% of primitives. Always L₀ — free.
- Reductive (N-to-1) — information concentrated. 17%.
- Transformative (N-to-N) — information rearranged. 56%.
- Generative (1-to-N) — information expanded. 4%.
This axis is where Landauer meets the stack. Any bijective operation can be implemented reversibly. Any reductive operation cannot. Information topology determines thermodynamic cost.
Why six
We tried fewer. Five axes is not enough — you can't distinguish between a sequential accumulator and a parallel reduction. Seven axes is redundant — you get duplicates. Six is the minimum number that uniquely positions every irreducible primitive, and it falls out of the structure, not the designer.
Every primitive in the stack has exactly six coordinates: family, abstraction level, thermodynamic cost, concurrency topology, state topology, and information topology. Two of these come from mathematics. Two come from computer science. Two come from physics. The intersection is the stack.