Performance Benchmarks

Q-Memory Platform Performance Benchmarks

This page provides detailed performance metrics for the Q-Memory photonic platform, organised by capability area. Figures marked as “projected” are engineering estimates based on published literature for comparable photonic systems; figures marked as “Phase 0 target” are validated through the Phase 0 fabrication run.

Optical Component Performance

Waveguide Propagation Loss

Silicon nitride waveguides are the primary photon-carrying medium. The target loss is:

Phase	Loss target	Significance
Phase 0	< 2 dB/cm	Validates process quality; acceptable for proof-of-concept
Phase 1+	< 0.5 dB/m	Required for fault-tolerant quantum operation

At 0.5 dB/m, a 10 cm photon path loses only 0.01% of its light — well within fault-tolerance budgets. Conventional silicon waveguides at ~2 dB/cm would lose far more over the same path, making them incompatible with fault-tolerant photonic quantum computing.

Phase Element Performance

Parameter	Phase 0 target	Phase 1 target
Power for π phase shift	< 25 mW	< 15 mW
Switching response time (thermal)	< 10 µs	< 10 µs
Switching response time (electro-optic)	Not available (Phase 0)	< 1 ns
Phase drift rate	< 5 mrad/min	< 2 mrad/min
Phase precision (non-volatile elements)	Not available (Phase 0)	Multi-level; > 5 bits effective

Beam Splitter Performance

Parameter	Target	Notes
Splitting ratio accuracy	±1% of 50:50	Validates lithographic process
Extinction ratio	> 25 dB (Phase 0)	Required for MZI interference quality
Insertion loss per element	< 1 mdB	Per 50:50 split
Fabrication variation	Characterised via test structures	Determines calibration requirements

Edge Coupler Performance

Parameter	Phase 0 target	Phase 1 target
Insertion loss per facet	< 2 dB	< 0.1 dB
Polarisation selectivity	> 20 dB TM suppression	> 30 dB
Wavelength range	1530–1570 nm	1530–1570 nm
Fibre array pitch	127 µm	127 µm (standard V-groove)

Quantum Operation Performance

Two-Photon Hong-Ou-Mandel Interference

The Hong-Ou-Mandel (HOM) visibility test measures how indistinguishable two photons are — a critical parameter for quantum gate fidelity. Two photons that are perfectly indistinguishable always bunch together at a 50:50 beam splitter; distinguishable photons do not.

Phase	HOM Visibility Target	Significance
Phase 0	> 90%	Minimum for useful quantum operations
Phase 1	> 99%	Required for fault-tolerant applications

HOM visibility below threshold means photons are too distinguishable to produce reliable quantum entanglement — making the quantum logic unreliable.

Quantum Gate Fidelity

For photonic quantum gates (fusion operations):

Gate type	Expected success probability	Fidelity (conditional on success)
Type-I fusion	25%	> 99% (limited by photon indistinguishability)
Type-II fusion	50%	> 98%
Single-qubit rotation (MZI)	100% (deterministic)	> 99.9% (limited by phase precision)

The probabilistic nature of two-photon fusion operations requires running multiple parallel copies of each operation and routing successful outcomes — implemented in the Phase 1+ architecture.

Real-Time Feed-Forward Latency

The feed-forward path — detector measurement to correction applied at a downstream element — is critical for measurement-based quantum computation:

Component	Latency target
Photon detection and amplification	~1 ns
ADC digitisation	~2 ns
FPGA logic (correction compute)	~5 ns
DAC output and element response (electro-optic)	~1 ns
Total feed-forward loop	~10 ns

A 10 ns feed-forward loop is sufficient for photonic quantum computing where photons travel at approximately 2 × 10⁸ m/s in the waveguide — the correction can be applied before the photon reaches the next stage if the optical path length exceeds approximately 2 m equivalent.

AI Acceleration Performance

Optical Matrix-Vector Multiplication Latency

The computation time for optical matrix multiplication is set by photon propagation time across the chip plus detector readout:

Matrix Dimension	Optical transit time	Total latency (incl. detector)	Memory-bound GPU (approx.)
4 × 4	< 1 ns	~10 ns	~10–100 ns
64 × 64	< 5 ns	~15 ns	~100 ns–10 µs
256 × 256	< 20 ns	~30 ns	~1–100 µs

Energy per Matrix Operation

Operation	Photonic platform	GPU (approx.)	Advantage
64×64 MVM (inference, fixed weights)	~1–10 nJ	~100–1000 nJ	10–100×
256×256 MVM (inference, fixed weights)	~5–50 nJ	~1–100 µJ	20–2000×

Note: The photonic advantage is largest when the same weight matrix is used many times. For single operations or rapidly-changing weights, the reprogramming overhead reduces the advantage.

Non-Volatile Weight Storage

For AI inference with fixed weights:

Scenario	Continuous phase control power	Non-volatile optical memory power	Saving
64-element network	~1 W	0 W	1 W
256-element network	~4 W	0 W	4 W
1024-element network	~15 W	0 W	15 W

This saving is structural — the weights are stored physically in the material state and require no power to maintain.

Photon Source Performance

Phase 0 (External Source)

Phase 0 uses an external fibre-coupled photon pair source. Typical parameters for the characterisation experiments:

Parameter	Value	Source
Centre wavelength	1550 nm	Telecom C-band
Bandwidth	~0.3 nm FWHM	KTP or PPLN SPDC source
Pair generation rate	~MHz at 1 mW pump	Typical for Phase 0 characterisation

Phase 1+ (On-Chip Source)

On-chip photon pair sources using spontaneous nonlinear optical processes in ring resonators:

Parameter	Target	Notes
Pair generation rate	> 10 MHz/mW pump power	Published literature for optimised rings
Photon purity (g²(0))	< 0.1	Needed for high-visibility quantum interference
Telecom wavelength	1550 nm ± 10 nm	C-band compatible

Summary: Phase-by-Phase Performance Targets

Capability	Phase 0	Phase 1	Phase 2
Optical mode count	4	~64	~256
Waveguide loss	< 2 dB/cm	< 0.5 dB/m	< 0.5 dB/m
HOM visibility	> 90%	> 99%	> 99%
Feed-forward latency	Not implemented	< 10 ns	< 10 ns
Max matrix dimension (AI)	4×4	64×64	256×256
Non-volatile optical memory	Not included	Included	Included
On-chip photon source	No (external)	Yes	Yes