29 out of 258
The Periodic Stack contains 258 computational primitives across 33 families. Of those 258, exactly 29 have dedicated silicon — hardware that implements the primitive directly, not through software emulation. That's 11% of the stack.
The other 229 primitives? They run in software, on general-purpose hardware that wastes orders of magnitude more energy than the thermodynamic minimum. They are the gaps. And the gaps are enormous.
The corridor is narrow
What makes this striking is which 29 primitives have silicon. They cluster in the same few families: linear algebra, reduction, activation functions, a handful of data-movement operations. The rest of the 33 families — 25 of them — have zero dedicated hardware.
No silicon for number-theoretic transforms. No silicon for formal verification primitives. No silicon for topological operations, spiking neural dynamics, symbolic rewriting, constraint propagation, or graph traversal at the hardware level. These families exist in the stack because the math demands them. The industry has simply never built chips for them.
Every chip lives in the corridor
Map each major chip architecture against the 258 primitives and you see the convergence:
- Apple M5 — ~31 primitives in hardware. CPU scalar ops, NEON/AMX vector ops, Neural Engine matmul and activation functions, GPU texture and rasterization units. Coverage: ~12% of the stack.
- Google TPU v5 — ~26 primitives. Systolic array matmul, reduction, softmax, a few activation functions, high-bandwidth transpose. Coverage: ~10%.
- NVIDIA H100 — ~28 primitives. Tensor cores for matmul and convolution, warp-level reduction, a small set of transcendentals (exp, log, rsqrt). Coverage: ~11%.
- AMD MI300X — ~25 primitives. Matrix cores, wavefront reduction, same transcendental set as NVIDIA. Coverage: ~10%.
- Groq LPU — ~11 primitives. Matmul, reduction, activation, transpose — stripped to the minimum for inference. Coverage: ~4%.
- Tenstorrent Wormhole — ~18 primitives. RISC-V scalar cores plus tensor operations, some sparse-matrix support. Coverage: ~7%.
Six different companies. Six different architectures. Six different transistor budgets. All of them land in the same narrow corridor of the stack — the linear algebra and activation families — because that's where the market demand is today: matrix-heavy deep learning inference and training.
25 families with zero hardware
The corridor accounts for roughly 8 of the 33 families. The remaining 25 families have no dedicated silicon anywhere in the industry. Some examples:
- Lattice-based cryptography — post-quantum security depends on lattice operations that run 100x slower in software than they would in a dedicated unit.
- Spiking neural dynamics — neuromorphic chips like Intel Loihi and SynSense Xylo cover a few primitives here, but most of the family is unaddressed.
- Symbolic rewriting — term rewriting, unification, pattern matching. The backbone of formal methods, proof assistants, and compiler optimization. All software.
- Constraint propagation — SAT solvers, SMT solvers, scheduling engines. Entire industries depend on these. No hardware.
- Topological operations — persistent homology, simplicial complex construction. Critical for scientific computing and emerging in ML. No hardware.
Each of these families contains multiple primitives with known thermodynamic bounds. The stack tells you exactly how much energy they should cost. The gap between that bound and the current software implementation is the engineering opportunity.
Why the corridor exists
Market gravity. Chip design is expensive. Companies build what sells. What sells is what the current software ecosystem demands. The current ecosystem demands matmul because deep learning dominates. So every chip vendor optimizes for matmul, and the corridor gets deeper but not wider.
This is rational behavior in the absence of a map. If you don't know what primitives exist outside the corridor, you can't evaluate whether to build them. The stack provides that map. It shows all 258 primitives, their costs, their dependencies, and their current silicon status. For the first time, you can see the full landscape — not just the 11% that the industry has explored.
The 89% opportunity
229 primitives with no dedicated hardware. 25 families with no silicon representation at all. Each gap is a chip that could exist but doesn't. Each gap is an energy reduction waiting to happen — often 100x to 10,000x compared to the software implementation.
The corridor is not a law of physics. It's a consequence of history and economics. The stack makes the full space visible. What gets built next is a choice.
Every modern chip covers roughly the same 11% of the Periodic Stack. 25 of 33 families have zero dedicated hardware. 229 of 258 primitives run only in software. The silicon corridor is narrow not because the math is narrow, but because the industry optimizes for today's workloads. The stack reveals the other 89%.