CatalogSystems
← systems
Frontier · Photonic

Silicon-photonic MAC tile

Matrix-multiply done in light at 1550 nm — sub-picojoule per MAC. The same descent, one paradigm out.

baseline: open silicon photonics — GDSFactory / SiEPIC components, AIM Photonics PDK
carrier 1550 nmenergy sub-pJ / MACPDK AIM Photonics (open MPW)core MZI mesh

The frontier exemplar, and the point where the classification flips. On the ring, the compute (the MCU) was reuse and only the sensor was the slice. Here the compute itself is the slice — you cannot buy a photonic matrix core off a reel; it is the thing openWafer designs, in the substrate where a multiply-accumulate costs the fewest joules. The passives, the laser, the readout, and the drive electronics around it are still reuse.

5 of 8 parts are reuse — 3 are the slice worth designing. That ratio is the NRE bill. The descent below is how you find it before you pay it.

Exploded

Decompose the product into its stack. Each layer is colour-coded by what it is to the next build.

1550 nm laser source (off-chip) Common · stable
Grating couplers — fiber I/O Common · stable
MZ modulators — input encode Opportunity
MZI mesh + phase shifters — the matrix core Opportunity
Ge photodetectors — readout Common · stable
CMOS TIA / DAC / ADC — drive Common · cyclic
Si / SiN waveguide substrate Common · stable

Bill of materials

stable ×4 cyclic ×1 Opportunity ×3

Every part is a real, sourced component from the open baseline — no invented part numbers.

MZI mesh (Clements/Reck) SiEPIC / GDSFactory components Opportunity · the slice openWafer designs
Matrix-multiply core — triangular/rectangular Mach-Zehnder mesh — the programmable linear-algebra fabric

This is the computer. There is no off-the-shelf photonic matrix core — the mesh topology, the loss budget, and the calibration are the design. The multiply-accumulate primitive realised in the substrate where it costs the least energy.

https://gdsfactory.github.io/gdsfactory/ →
Thermo-optic phase shifter PDK standard cell Opportunity · the slice openWafer designs
Weight tuning — heater-tuned π phase shift per MZI arm

Thermal phase shifters burn static power continuously — the scaling blocker. The slice openWafer designs: non-volatile phase shifters (phase-change / MEMS / BTO) that hold weights at zero idle power.

Mach-Zehnder modulator AIM Photonics PDK Opportunity · the slice openWafer designs
Input encode — electro-optic intensity modulator @ 1550 nm

Compact slow-light EO modulators (TeMPO-class) cut the area and drive energy of the encoder — co-designed with the CMOS driver below.

https://www.aimphotonics.com/pdk →
Grating coupler AIM / SiEPIC standard Common · stable · reuse as-is — zero NRE
Optical I/O — fiber-to-chip coupler, ~1550 nm band

Mature PDK passive. Reuse as-is.

https://github.com/SiEPIC →
Germanium photodetector AIM Photonics PDK Common · stable · reuse as-is — zero NRE
Readout — on-chip Ge PD, 15 Gbps tested; squared modulus gives the nonlinearity

Standard, qualified PDK component. Reuse.

https://www.aimphotonics.com/pdk →
1550 nm DFB laser commodity telecom Common · stable · reuse as-is — zero NRE
Light source — off-chip continuous-wave carrier

Telecom-volume commodity. Reuse.

CMOS TIA + DAC/ADC standard mixed-signal Common · cyclic · refreshes each generation — reuse the current one
Electronic drive — transimpedance readout + modulator drivers + converters

Standard CMOS that refreshes on a node clock — reuse the current generation; the co-design opportunity is the interface, not the process.

Si / SiN waveguides PDK passives Common · stable · reuse as-is — zero NRE
Substrate — low-loss routing on SOI / silicon nitride

Settled PDK passive layer. Reuse.

The slice openWafer designs

On the ring, the compute was reuse — an off-the-shelf MCU — and the slice was the sensor. Here it inverts: the compute is the slice. The MZI mesh, the phase-shifter technology, and the modulator are the matrix-multiply primitive realised in light, in the substrate where it dissipates the fewest joules per operation. The passives, the laser, the photodetector, and the CMOS drive are all reuse. openWafer’s design is the computational core itself — and specifically the move that makes photonic compute scale: replacing power-hungry thermo-optic tuning with non-volatile phase shifters, so the mesh holds its weights at zero idle power. That is design-from-math, pick-the-substrate: put the matrix-multiply where its friction is lowest, and design only the part no one can hand you.

Open sources — decomposed and linked, not hosted