Introduction

TALON IR is a neuromorphic intermediate representation designed for Type 1 Compute hardware. The ecosystem consists of seven packages that work together to bridge ML frameworks with neuromorphic deployment, hardware mapping, FPGA compilation, and real-time event streaming.

High-Level Architecture Overview

TALON

For most users, the recommended entry point is TALON — Tactical AI at Low-power On-device Nodes:

pip install t1c-talon

This installs all ecosystem packages and provides both the talon and t1c CLI commands. See the TALON Documentation for details.

Package Overview

Package	PyPI Name	Purpose	Key Dependencies
talon.sdk	t1c-talon	SDK meta-package and CLI (talon / t1c)	all packages below
talon.ir	talon-ir	Core IR primitives and HDF5 serialization	numpy, h5py, rustworkx
talon.bridge	talon-bridge	PyTorch export/import bridge	talon-ir, torch, snntorch
talon.viz	talon-viz	Interactive graph and spike visualization	talon-ir, pillow
talon.graph	talon-graph	Graph partitioning, placement, routing, resource allocation	talon-ir, rustworkx
talon.backend	talon-backend	Backend compilation, CPU simulation, HLS4ML FPGA backend	talon-ir, talon-graph, numpy
talon.io	talon-io	Event streaming, format conversion, neural encoding	numpy

Talon Modules/Packages

Design Philosophy

The architecture mirrors the ONNX ecosystem:

PyTorch Model → bridge.to_ir() → TALON IR Graph → ir.write() → .t1c file
                                                     │
                                        ┌────────────┼────────────┐
                                        ▼            ▼            ▼
                                   graph.partition  backend.compile  viz.visualize
                                   graph.route      backend.simulate
                                   graph.place      backend.profile
                                        │            │
                                        ▼            ▼
                              Hardware Mapping   HW Descriptor (JSON)

(This assumes from talon import ir, bridge, graph, backend, viz.)

This separation enables:

• ML engineers export models without understanding hardware details
• Hardware teams consume standardized IR without touching PyTorch
• Deployment engineers partition, map, and compile graphs to hardware targets
• Visualization works independently of all other packages
• I/O engineers stream and encode real-time event camera data

Key Features

• 36 primitives covering SNNs, ANNs, hybrid architectures, Ghost blocks, and detection heads
• Hybrid ANN-SNN support with activations (ReLU, Sigmoid, GELU, Softmax), normalization (BatchNorm, LayerNorm), and dropout
• NIR-compliant LIF neuron dynamics for interoperability
• RustworkX-powered graph algorithms for 100x faster validation and cycle detection
• HDF5 serialization for efficient storage and portability
• snnTorch integration via thin wrapper functions
• Graph partitioning across neuromorphic cores (greedy, edge-map, spectral algorithms)
• Core placement and spike routing on physical mesh topologies
• Backend compilation to hardware descriptors (JSON/binary) for runtime configuration
• CPU simulation and profiling with from-scratch latency and energy estimation
• HLS4ML FPGA backend for Xilinx Zynq targets with C++ header generation
• Netron-style visualization with clickable nodes, pattern detection, and parameter inspection
• Spike/event visualization for neuromorphic datasets (23-30M events/sec)
• Event I/O with vectorized EVT2/EVT3 decoding (>5M events/sec), AEDAT4 unpacking, UDP streaming

Use Cases

TALON IR is designed for:

• Model deployment to Type 1 Compute neuromorphic hardware
• Hybrid ANN-SNN architectures with encoder-SNN-decoder patterns
• FPGA prototyping via HLS4ML config generation and Vivado TCL scripts
• Hardware mapping partitioning graphs across neuromorphic cores with placement and routing
• Model exchange between ML and hardware teams
• Architecture visualization for debugging and documentation
• Event-based data processing real-time event camera streaming and neural encoding
• Round-trip verification ensuring export/import fidelity
• Energy and latency profiling from-scratch CPU simulation with MAC-based energy estimation

Project Repositories

Repositories (Git clone names) and their installed package names:

• t1ctalon: SDK meta-package and CLI (installs as t1c-talon; commands: talon and t1c)
• talonir: Core IR package (installs as talon-ir; use from talon import ir)
• talonbridge: PyTorch bridge (installs as talon-bridge; use from talon import bridge)
• talonviz: Visualization tools (installs as talon-viz; use from talon import viz)
• talongraph: Graph partitioning and hardware mapping (installs as talon-graph; use from talon import graph)
• talonbackend: Backend compilation and simulation (installs as talon-backend; use from talon import backend)
• talonio: Event streaming and I/O (installs as talon-io; use from talon import io)

Next Steps

• Quick Start — Export your first model in 5 minutes
• Installation Guide — Set up all packages
• Architecture Overview — How the packages work together
• TALON Overview — Unified entry point and CLI

Tutorials

• TALON IR Basics — Primitives, graphs, serialization
• talonviz — Graph visualization and event processing
• TALON SDK — Analysis, profiling, linting, fingerprinting
• Bridge — Export/import PyTorch models
• Ghost & Detection — GhostNet and detection pipeline
• Hardware Mapping — Partition, route, place
• Backend — Simulate, profile, FPGA
• Event I/O — Neural encoding, HDF5, throughput
• snnTorch Integration — snnTorch export/import round-trip
• End-to-End Pipeline — Full workflow: model → export → analyze → partition → simulate → profile