Phase-Change Photonic Memory

The HBM lesson taught modern AI hardware's ugliest secret: the arithmetic is nearly free, and the energy bill is dominated by hauling the weights — picojoules per byte shuttled from DRAM stacks, thousands of times the cost of the multiply they feed. Photonic accelerators do not escape this. A microring weight bank or an MZI mesh must be told its weights by electronics, and must burn static power (in heaters, typically) every second it holds them. Switch off the power and the entire network amnesias.

Now for the trick, borrowed from the least glamorous corner of the optical industry. Deposit a speck of GST — the chalcogenide glass Ge₂Sb₂Te₅, the coating inside every rewritable DVD — a few hundred nanometres long on top of a silicon waveguide. GST has two stable solid phases with wildly different optical absorption, and it will sit in either one, unpowered, for years. The transmission of the waveguide underneath becomes a non-volatile optical weight: written with a light pulse, held for free, and — the part that changes the architecture — read by the very act of computing. Light carrying an activation passes through the cell and emerges multiplied by the stored weight. The memory is the multiplier. Nothing is fetched, because the data never had to move.

Two phases, two transmissions

GST's two personalities come from atomic order. In the crystalline phase the atoms sit on a lattice with resonant bonds that soak up telecom-band light: high refractive index, high extinction coefficient, strong absorption. In the amorphous phase the lattice is a frozen jumble, the resonant bonding is broken, and the material turns nearly transparent. Both phases are stable at room temperature — there is an energy barrier between them, so neither decays into the other on any timescale you care about.

	Amorphous (quenched)	Crystalline (annealed)
atomic structure	disordered glass	ordered lattice
optical absorption at 1550 nm	low (≈ 0.3 dB/µm here)	high (≈ 6 dB/µm here)
cell transmission (1 µm cell)	≈ 0.93	≈ 0.25
as a stored weight	w near 1	w near 0
how to reach it	melt-quench: short intense pulse	anneal: longer moderate pulse

Switching between them is cookery. To amorphise (erase toward transparency), hit the cell with a short, intense pulse — nanoseconds, enough to melt the GST past ~900 K — and let it cool so fast (>1 K/ns, automatic at this size) that the atoms freeze in place before they can find the lattice: melt-quench. To crystallise (write toward absorption), apply a gentler, longer pulse that holds the cell above the crystallisation temperature (~430 K) but below melting, giving the atoms time to snap onto the lattice. Both pulses can be sent down the same waveguide that carries the data — or applied electrically through a microheater. And because crystallisation can be stopped partway, the cell is naturally multi-level: a partial crystalline fraction gives an intermediate transmission, and experiments have resolved dozens of distinguishable levels — five to six bits in one cell.

Compute where the weights are

A single cell multiplies: P_{\text{out}} = w \cdot P_{\text{in}}, with the weight w stored in the atoms themselves. Tile cells into a crossbar — input activations encoded as optical powers on N wavelengths, a GST cell at each crossing, photodetectors summing each output column — and one pass of light performs N \times M multiply–accumulates with zero weight-fetch traffic. This is in-memory photonic computing: the architectural answer to the memory wall is to abolish the distinction between memory and processor. It is not a paper fantasy — a 2021 Oxford–Münster–Exeter collaboration (Feldmann et al.) ran convolutions through a GST crossbar driven by a frequency comb at trillions of MACs per second, an experiment this module will meet again two lessons from now.

The economics are lopsided in an interesting way. Holding a weight costs nothing — contrast that with the microring bank's heater or DRAM's refresh. Reading a weight costs nothing beyond the light you were sending anyway. Writing a weight is the expensive act: hundreds of picojoules to nanojoules of pulse energy, over nanoseconds to microseconds. Amortise it:

const tAmorphous = 0.93, tCrystalline = 0.25; // 1 µm cell transmissions const contrastDb = 10 * Math.log10(tAmorphous / tCrystalline); console.log("contrast: " + contrastDb.toFixed(2) + " dB"); const levels = 32; // resolvable transmission levels console.log("bits per cell: " + Math.log2(levels)); // write once, read forever: amortised write cost per inference const writeEnergy = 5e-10; // 0.5 nJ per weight update for (const inferences of [1, 1e3, 1e6, 1e9]) { const perInference = writeEnergy / inferences; console.log(inferences + " inferences → " + perInference.toExponential(1) + " J per inference"); }

Half a nanojoule looks terrible next to a femtojoule MAC — until you divide by the number of times the weight is used. A deployed inference accelerator writes its weights once and reads them billions of times, driving the amortised write cost to attojoules. The same arithmetic read backwards is the technology's confession: any workload that updates weights constantly — training — pays the full write price on every step, and phase-change memory is the wrong tool for it. Non-volatile photonic weights are an inference technology, by construction.

Phase-change materials were championed from the 1960s by Stanford Ovshinsky — a self-taught inventor without a university degree who founded a company on the idea that disordered materials could switch and remember. The industry first bit in optics: rewritable CDs and DVDs store bits as amorphous marks melt-quenched into a spinning crystalline GST film by the drive's laser, read out as reflectivity changes — billions of discs, the largest deployment of phase-change physics in history. Electronics followed with 3D XPoint / Optane memory, storing bits as GST resistance instead. The photonic-memory idea (first integrated demonstration: Ríos et al., 2015) closes a satisfying loop: the disc's optics and the memory chip's non-volatility, reunited on a silicon waveguide — the same melt-quench trick at its third career, this time holding neural weights rather than home movies.

Three pieces of fine print keep GST weights honest. Drift: the amorphous phase is a glass, not an equilibrium state; it slowly relaxes structurally, and the cell's optical properties creep logarithmically in time. Finely spaced levels smear into their neighbours, which is why "6 bits" is a lab number and deployed systems budget fewer bits plus periodic recalibration. Wear: each melt-quench cycle stresses the film (think repeated recasting of the same metal part); endurance of 10^{8}–10^{12} cycles is superb for a weight rewritten occasionally and hopeless next to DRAM's effectively unlimited writes — another reason this is inference hardware. Asymmetry and sign: writes are slow (ns–µs) and energetic while reads are fast and free, and a transmission w \in [0, 1] is amplitude-only and non-negative — negative synaptic weights need a differential pair of cells with the detector subtracting, doubling the cell count. None of these is fatal; all of them are why the demonstrated systems are small crossbars, while wafer-scale in-memory photonic tensor engines remain a roadmap item.

Where this goes next

This lesson fixed the weights into the fabric of the circuit. The next lesson asks the complementary question: what if the circuit itself were not fixed? Take a featureless lattice of tunable couplers and program it — after fabrication, in software — to become a filter, a delay line, an interferometer, or all three at once: the FPGA of light.