nx_gate_energy_test.nx

nx_gate_energy_test.nx -- the EFFICIENCY leg of the triangle, measured BITS-UP from the silicon we built, not borrowed from an OS. This is the PORTABLE layer: it runs identically on ANY hardware because it owes nothing to the host (no RAPL, no /sys, no privilege) -- the energy is a property of the GATE DESIGN. THE GOOD TRIANGLE (operator's framing): SPEED = critical-path gate DEPTH (signals ripple through the carry chain) SIZE/AREA = gate COUNT (how much silicon the design occupies) EFFICIENCY = switching ACTIVITY = gate-output TOGGLES per op. Dynamic energy in real silicon is ~ (toggles * C * V^2): every gate output that flips charges/discharges a capacitance and burns joules. Counting toggles is exactly how chip designers estimate dynamic power. CAPABILITY = 1:1 CORRECTNESS (the sum is right for every vector). CITATIONS (Cardinal #4 -- real sources): the dynamic-power law P = a*C*V^2*f and switching-activity-based estimation are textbook -- Weste & Harris, "CMOS VLSI Design" 4e; Najm, "A Survey of Power Estimation Techniques in VLSI Circuits," IEEE Trans. VLSI Systems, 1994. Ripple-carry critical-path ~2W gate delays is standard (Weste & Harris, adder chapter). The energy/time/memory tradeoff this triangle measures is empirically grounded in Pereira et al., "Energy Efficiency across Programming Languages," ACM SLE 2017 (the benchmarks-game-based study). We build a REAL ripple-carry adder out of PRIMITIVE 1-bit gates (XOR/AND/OR), simulate it on consecutive input vectors, and count how many gate outputs flip. Key physical truth this proves: ENERGY IS DATA-DEPENDENT. The same correct adder burns more on random inputs (carries propagate, ~half the bits flip) than on small increments (short carries, few flips). That is WHY a garden sensor must MEASURE its real workload's energy, not assume it -- the companion real-hardware probe (nx_energy_probe) reads the node's actual power sensor and calibrates THIS toggle count into joules for whatever silicon it landed on. Known answer: correctness holds for all vectors AND random-workload toggles > increment-workload toggles (data-dependent energy). exit 0.