Not all gaps are equal
229 primitives lack dedicated silicon. If you could only build hardware for one of them, which one should it be? The answer isn't "the most important one" — importance is subjective. The answer is the one with the most downstream dependents. Build it and the most other things get faster.
The out-degree leaderboard
Count how many composition edges leave each primitive. That's its out-degree — the number of other operations that are algorithmically built from it. The top 4 that already have silicon (#18 matmul, #21 reduction, #4 FMA, #15 mod-exp) are the reason the entire AI/HPC industry converges where it does.
The next 4 that don't have silicon are the highest-leverage targets:
- #76 Matrix decomposition — 11 downstream. QR/LU/SVD. No chip.
- #94 Differentiation — 11 downstream. Every gradient, every PDE. No chip.
- #59 Symbolic term rewriting — 10 downstream. Gröbner, analytic continuation, proof search. Software only.
- #95 Integration / quadrature — 9 downstream. Every FEM, every Feynman integral. No chip.
The L₀ leverage list
36 primitives in the stack are thermodynamically free (L₀ / bijective). 6 already have silicon (#2 bitwise, #12 NTT, #23 FFT, #40 scatter/gather, #41 embedding lookup, #69 entropy coding). The other 30 run on energy-burning general-purpose hardware.
The energy-savings multiplier for an L₀ primitive is out_degree × invocations_per_call × energy_per_invocation. For FFT (out-degree 8), that multiplier is enormous. For Möbius/zeta transform (out-degree 3), it's smaller but the per-invocation savings are total — from L₂ cost down to L₀ floor.
The in-degree view
The in-degree tells you the opposite story: which operations are the most algorithmically composite — built from the most sub-operations. These are the highest-value single-chip targets because one piece of silicon concentrates the most algorithmic work.
#229 HMC lattice gauge sweep (in-degree 3), #180 ZK proof generation (in-degree 3), #274 lattice Boltzmann (in-degree 3) are the most composite operations in the stack. Each would need a chip that fuses multiple sub-operations into a single datapath.
The leverage leaderboard turns the 229 hardware gaps into a ranked priority list. The top 4 gaps by out-degree (#76, #94, #59, #95) are the primitives where silicon investment multiplies across the most downstream operations. The top L₀ gaps are where the energy savings multiply. Open the instrument → Section V to see the full ranking.