PHOTONICS

ULTRA-LOW POWER PHOTONIC TENSOR CORE

Technical Approach

Innovation: a chip-scale, energy efficient (1fJ/op), reconfigurable, low latencytensor core, with roadmap to volume manufacturing

Enabling technology: novel non-volatile EO switch (FLOW) à photonic tensor core (PTC) w/ zero-power state maintenance

Enabled Capability: Near-zero-energy inference for imagery and/or RF signal classification

Impact: Deep learning classifiers and processing integrated into power-constrained UAS edge platforms for 10X longer duration passive surveillance capabilities

Additional impact in processing parallel linear operations such as DFT

Accomplishments/Demos:

>72 hour state retention of new FLOW optical switch
Emulated CNN demo of FLOW enabled PTC
Rack mount chassis RF-in/RF-out to demo unit performance
Hardware demo of PTC enabled CNN for image classification (by 10/23)

EXPECT FLOW ENABLED PTC ARCHITECTURES TO SUPPORT LOW
POWER & LATENCY COMPUTE FOR EDGE PLATFORMS

SWAP constraints in edge platforms allocate as little as 1% of sensor volume to processing resources and demand very low power operation.

Applications: UAS image processing energy reduction. Future: Ultra-low-power AI spaces (mobile, handset, commercial drone, etc.). Expand beyond image classification into RF, EW, spectrum optimization, and others.

FUTURE OF FLOW-BASED PHOTONIC TENSOR CORE

Edge AI Target Applications:

EO/IR Image Classification
Wideband RF Spectrum Sensing and Signal Classification
Multi-modal EW & Spectrum Superiority
Autonomous Navigation in Denied Areas

FLOW-based AI systems offer unparalleled operating bandwidth, latency, SNR, and power over electronics Significant power savings versus state-of-the-art optical neural networks, enabling UAS deployment.

Low-power optical processing hardware enhances capabilities in broad range of small form factor platforms

UAS processing energy reduction of 10X

Offloading of processing
Longer UAS durations
Surveillance from increased distance
Frees up processing capability for non-ML sensor processing

Vision: Full parallel system capability – tiled PTCs

Low Latency
Large instantaneous and operating bandwidth
High SNR
Low power
Small SWAP

USING A SINGLE 9X9 PROTOTYPE TO DEMONSTRATE A CONVULUTIONAL
NETWORK FOR CLASSIFICATION OF SURVEILLANCE TARGETS

Modular network design using optical PTC hardware as building block for linear operations, nonlinear operations performed in digital domain.

Examining direct implementation through generalized matrix multiplication and compression of large networks to accommodate small prototype core

Modeling on MNIST data set suggests only a few percent reduction in classification when comparing PTC enabled CNN to all digital CNN.

Surveillance target classification evaluation in progress.

KEY DIFFERENTIATORS & COMPETITIVE ADVANTAGES

OPTIMIZED FOR ANALOG & SENSING SOLUTIONS

Platform & internal know-how provides a functioning bedrock for fast-prototyping

Fast transition from prototyping to volume production

Mature Technology Development from self-interference calculations and signal extraction

UNMATCHED FUNCTIONALITY PER CHIP

Innovatively pack high number of functional components onto each retical, resulting in a high performance, multi-use chips

Profound expertise resulting in ultra low loss (<0.4dB/m) & noise output

ENTERPRISE-GRADE PDK LIBRARY

Module-level process optimization at every step of the process

Built on standard Soitec SOIs for ease of production

Internal know-how out of Motorola and Intel fabs result in best-in-class foundry processes

ADVANCED 3D PHOTONICS PLATFORMS: SiN + c-SI

GXC Photonics combines high-performance Ultra Low Loss SiN waveguides with the rich capabilities of crystalline silicon in a monolithic platform enabling high fidelity applications

Combines Si waveguide layer (2.2dB/cm) with an
ultra-low-loss Si₃N₄ waveguide layer (0.016dB/cm)
GXC Platform when compared to Si₃N₄ -only
platforms:

> 10X speed improvement (<25μsec)
> 10X die size reduction (30μm bends)
> 7X reduced power consumption (47mW @ Vπ)

GXC PHOTONICS’ UNIQUE SOLUTION ADDRESSES GLOBAL PROBLEMS

PROBLEM

System OEMs must work with multiple agents to satisfy demands of products they wish
to commercialize

Wafer manufacturers take many months to deliver prototypes

Presented system designs are difficult to instantiate efficiently, cost effectively and within acceptable foundry yields

Few teams exist capable of designing & delivering photonic-based systems

GXC PHOTONICS SOLUTION

End-to-end turnkey system delivery at pace

Smaller, efficient, feature-rich units

Technology vertically integrated by a single provider

Rapid prototyping drives faster time-to-production

INTERNAL CAPABILITY PROVIDES CUSTOMERS WITH TURNKEY SOLUTION

Unrivalled Vertical Integration from Prototyping to Fab Transfer

PROCESS & DEVICE INNOVATION

Defense-focused
World-class product engineers & PIC designers
Adaptable & tunable GXC Photonics Platform
System specification

DESIGN & PIC MANUFACTURING

Uniquely combine RF, Optics & Mechanical components
Rapid prototyping in 8-12 weeks vs Industry 40-65 weeks
Designed to accommodate efficient fab processes and high-yielding techniques

ADVANCED PACKAGING

Multi-chip 3D packaging
RF compatible
Ultra low loss optical I/Os
Low noise
Optimized for linear & non-linear
elements

SYSTEM INTEGRATION & TESTING

Validated transfer to production capability
Formally scoped & validated by TSI Semi & SkyWater
Skilled team capable of integrating entire modules with multiple
functionalities
Turnkey operation