How llandsmeer/iotf works
Top Layer: Model Definition
Middle Layer: Runner Framework
Bottom Layer: Hardware Execution Layer
Component Details
a) Model Definition Layer:
• Subcomponents for:
• Initial neuron state generation
• ONNX model creation
• TensorFlow function creation
• 3D connection sampling
• Timestep handling
b) Runner Framework Layer (Core):
• Branch out to show the different runner implementations:
• ONNX Runners (CPU, CUDA, TensorRT)
• TensorFlow Runners (CPU, TPU)
• Hardware-Specific Runners (Graphcore, GroqChip)
c) Hardware Execution Layer:
• CPU
• CUDA GPUs
• TensorRT
• Graphcore IPUs
• Groq GroqChips
• TPUs
Data Flow Indicators
• Model compilation flow (Model → Runners)
• Execution paths (Runners → Hardware)
• Benchmark data collection paths
• Test validation flows
Detailed Elements
• The plugin architecture for different hardware backends
• The optimization variants (especially for GroqChip runners)
• The separation between base implementations and hardware-specific optimizations
• The benchmark data collection and reporting system
Notes Section
A notes section indicating:
• Hardware requirements
• Optional components (based on available hardware)
• Key performance considerations
• Testing requirements
• Hierarchical component representation
• Clear inheritance and dependency visualization
• Export to high-resolution formats
Remember to emphasize the benchmark nature of the project and how it enables performance comparison across different AI acceleration platforms for neural simulation workloads.