Step 1.1: Define workload set
Select 3 to 5 representative persistent workloads (robotic perception, long-context sequence tasks, edge sensor streams).
Implementation roadmap
Detailed roadmap for PiP research and system implementation
This roadmap defines the full execution path from simulation foundations to benchmark publication. Every phase includes objectives, concrete tasks, required tools, artifacts to produce, and phase exit criteria.
Execution model
Work proceeds in strict loops: design, simulate, fabricate, measure, calibrate, benchmark. No downstream phase starts without upstream gate evidence.
Mechanism
Define objective to build artifact to run gate suite to decide advance or iterate.
Comparison mode
Left panel: execution loops repeatedly move parameters and context through memory hierarchies.
Right panel: inference appears as propagation across a learned physical operator with direct state readout.
FlowStateMeasurementConstraint
Phase map
The timeline below is the master rollout map. Each phase has a dedicated section with detailed implementation instructions.
Stage 1
Physical operator validation
Validate operator fidelity across benchmarked transforms and stability windows.
ExitDeterministic mapping and rank threshold confirmed
Statusin validation
Stage 2
Programmable fabrication
Establish repeatable fabrication pathways for tunable and task-specific structures.
ExitFabrication variance remains within acceptable tolerance band
Statusplanned
Stage 3
Developer toolchain
Expose compiler interfaces for mapping model structure to physical geometry.
ExitCompiler to runtime handoff validated across test suites
Statusplanned
Stage 4
Ecosystem deployment
Integrate substrate, readout, and software interfaces into deployable systems.
ExitEnd-to-end integration passes persistent workload trials
Statusplanned
Phase 1: Foundations and reproducibility setup
Step 1.2: Freeze baseline metrics
Record current GPU/NPU throughput, energy per update, latency distribution, and quality targets using identical datasets.
Step 1.3: Build reproducibility harness
Create run manifests, seed controls, environment fingerprints, and artifact checksums for all benchmark jobs.
Step 1.4: Define acceptance thresholds
Set pass criteria per workload and specify non-negotiable quality floors before any performance claims.
Phase 2: Operator design and simulation convergence
Build the compile-to-physics path and converge operator behavior in differentiable simulation before committing to fabrication.
Training complexity remains at design time. Fabricated operators contain learned mappings without retaining datasets or digital weight files at inference time.
FlowStateMeasurementConstraint
Step 2.1: Graph lowering
Transform selected model subgraphs into physical operator candidates with explicit geometry constraints.
Step 2.2: Simulation optimization
Run geometry optimization sweeps and track convergence of mapping fidelity, rank, and separability.
Step 2.3: Tolerance stress tests
Apply perturbation simulations for fabrication drift and thermal variance to estimate deployment robustness.
Step 2.4: Pre-fabrication signoff
Only promote candidates that satisfy gate suite across all representative workloads.
Phase 3: Fabrication readiness and metrology
Step 3.1: Layout export and process package
Generate mask-ready layouts, process recipes, and lot-level metadata templates.
Step 3.2: Pilot lot execution
Fabricate pilot sets and collect metrology traces for geometric fidelity and optical response.
Step 3.3: Calibration model build
Fit readout calibration models, compare against simulation predictions, and quantify error residuals.
Step 3.4: Fabrication gate decision
Advance only if variance remains within gate thresholds for all required operator classes.
Phase 4: Runtime integration and deployment path
Integrate calibrated operators into hybrid runtime services so developers can deploy with standard model-to-runtime workflows.
01
Input representation
A signal is encoded as optical boundary conditions derived from model input.
02
Physical operator
Engineered photonic geometry performs transformation through propagation.
03
Measured state
Output state is sampled at designated readout points with calibrated sensing.
04
Output
Measured response is decoded into task-specific prediction space.
Focused stage
A signal is encoded as optical boundary conditions derived from model input.
FlowStateMeasurementConstraint
Step 4.1: Runtime adapters
Implement adapters for compile artifacts, hardware mappings, and runtime invocation contracts.
Step 4.2: Observability and tracing
Emit per-run telemetry for state continuity, latency, energy traces, and calibration confidence scores.
Step 4.3: Edge envelope testing
Run sustained workloads under constrained power and thermal envelopes to confirm stability.
Step 4.4: Integration freeze
Freeze runtime interfaces only after compatibility passes across all target integration paths.
Phase 5: Benchmark publication and promise tracking
Step 5.1: rerun benchmark matrix with final calibrated operators and fixed baselines.
Step 5.2: generate reproducibility package (configs, datasets, seeds, artifact hashes).
Step 5.3: publish promise deltas against GPU/NPU baselines with quality parity evidence.
Step 5.4: open the next revision loop in formal specifications based on measured outcomes.
Required tools by workstream
ML and compiler: PyTorch or JAX, graph transform passes, artifact packaging.
Simulation: FDTD/FEM class solvers, optimization sweeps, tolerance perturbation engine.
Fabrication: layout toolchain, process control records, lot traceability and metrology capture.
Runtime: calibration service, hybrid execution runtime, telemetry and benchmark orchestration.
Execution instructions
Instruction 01: start from formal specification version and pin all toolchain versions.
Instruction 02: execute phase tasks in order and log all produced artifacts with checksums.
Instruction 03: run gate suite at each phase boundary; do not bypass failed gates.
Instruction 04: update specification with measured deltas before entering the next phase.
The binding contract for updates is maintained in Formal Specifications.
Risk controls and phase exits
Model risk
If quality drifts below threshold, rollback operator candidate and rerun simulation convergence cycle.
Fabrication risk
If lot variance exceeds tolerance, freeze deployment path and issue revised process recipe before retry.
Runtime risk
If continuity or latency regress, keep previous stable runtime adapter as active release baseline.
Benchmark integrity risk
If reproducibility package is incomplete, claims cannot be published and must remain provisional.