ADAMAS

Each bar says how many times better than silicon a material is at one job. Diamond's bars are the tallest, sometimes by tens of thousands.

BFOM = εμE_c³ scores conduction loss in a power switch; BHFFOM = μE_c² switching loss; JFOM the power–frequency product; KFOM the heat-limited switching density. The log axis hides how extreme diamond's 10 MV/cm breakdown field is: it enters cubed.

Values are room-temperature single-crystal numbers with diamond mobilities at the ultrapure upper bound; the figures assume full dopant ionization, which fails in diamond (panel 3), so read them as ceilings and pair them with panel 2's measured points.

Every line is the best a material could ever do. The stars are what people have actually built with diamond. They already beat silicon's best-possible line.

R_on,sp = 4BV²/(εμE_c³) for a one-sided drift region. The measured 2022 device beats ideal silicon by about 90× and trails ideal SiC; a vertical diamond device at 1% of its own limit would match GaN.

Measured points are lateral hydrogen-terminated NO₂-doped MOSFETs on heteroepitaxial wafers; the 3659 V device's R_on is derived from its reported Baliga figure of merit. SiC's T^2.4 mobility law is applied to its limit line; diamond's bulk-doped R_on falls with temperature instead (chapter E9).

Adding impurity atoms is how you make a semiconductor conduct. In diamond most of those atoms just sit there at room temperature. Heat wakes them up.

Charge neutrality with a single acceptor level E_A: p(p+N_D)/(N_A−N_D−p) = (N_V/g)e^−E_A/kT. Boron at 0.37 eV and phosphorus at 0.57 eV are 8 to 13 kT deep at 300 K.

Effective mass 0.8 m₀, degeneracy 4, no compensation. Above about 3×10²⁰ cm⁻³ the boron band merges with the valence band and the model no longer applies (metallic conduction, superconductivity). Surface transfer doping sidesteps activation entirely (chapter 3).

The NV center glows red. When a microwave hits exactly its note (2.87 GHz), the glow dims a little. A magnet splits that note into two, and the crystal has four kinds of NV pointing different ways, so you can see up to eight dips.

H/h = D(S_z² − 2/3) + γ_eB·S with D = 2.870 GHz and γ_e = 28 MHz/mT. Each of the four ⟨111⟩ orientations sees a different projection B_∥, giving lines at D ± γ_eB_∥.

Second-order perpendicular shift 3γ²B_⊥²/(2D) included; hyperfine structure (2.16 MHz for ¹⁴N) omitted at these line widths. Contrast fixed at 7.5% per orientation for an ensemble. The Python package includes the hyperfine term.

Two qubits feel each other through magnetism, but only when very close. The surprise: even with perfect memory, the gate fails if the qubit was not set up right at the start (that is the q slider).

ν = 52 MHz·nm³/r³; gate time 1/(2ν); with a spin echo the coherence factor is w = exp(−(t/T₂)³) and the Bell fidelity is q²(1+w)²/4 + (1−q²)/4.

With the published pair's parameters (25 nm, 19.7 kHz double-quantum coupling, T₂ 150 and 514 µs) the coherence-limited fidelity is 0.998; the measured 0.67 corresponds to q ≈ 0.75, the NV⁻ charge fraction under green light. See chapter E5, sections E5.1b and E5.1c.

A power switch wastes energy two ways: resistance while on, and charging its own capacitance every time it flips. The best material wastes the least of both.

P_min = 2IV√(fR_on,spC_oss,sp/2) with C_oss,sp ≈ εE_c/(2BV). Gate charge, reverse recovery, and GaN dynamic R_on are not modeled.

Hard-switched half-bridge, duty 0.5, area-optimized per switch. The measured diamond curve scales the 2022 device's R_on,sp as BV² to the chosen voltage class. Ideal diamond is a ceiling; today's device beats ideal silicon and trails ideal SiC. Chapter E9.

Quantum computers fix their own mistakes by spreading one answer over many qubits, like voting. That only works if each qubit and each connection is already pretty good. Green squares are where adding more qubits makes things better.

Planar surface code, phenomenological noise, minimum-weight perfect matching. Data error per round = in-cell error + 4·(8/15)·link error; measurement error = readout error + the same link term. Break-even (black dashed line): link error of about 1.3% per gate.

2000 shots per point, so entries below about 5×10⁻⁴ are upper limits and the green corner is noisy. The mapping from hardware rates to the two phenomenological rates is first order; hook errors and slow-readout idling need the circuit-level model X-1b. Validation: equal data and readout error crosses at 3.1% versus the published 2.93% (chapter E6).

Even a giant quantum computer would fit on a fingernail of diamond. The problem is speed: reading a diamond qubit takes about a thousand times longer than a superconducting one, so big jobs take years instead of hours.

Physical qubits from published surface-code estimates; 4 qubits per NV cell; time scales linearly with the cycle. Area is never the constraint; readout time and yield are.

Wall-clock times assume the published qubit counts at 10⁻³ physical error and a 1 µs reference cycle. Electrical readout (proposed, 100 µs) would bring a year down to a month. Chapter E10.