Exploring Coarse-Grained Reconfigurable Architectures (CGRA) and building a CGLA ("LA" here is Linear Array) based AI accelerator prototype on FPGA.
This project explores the design principles, architecture, programming models, and implementation challenges of CGRAs, with a particular focus on low-power AI acceleration.
The long-term goal is to develop a small experimental CGRA architecture and implement a CGLA prototype on FPGA.
The project will progress from architectural research to RTL implementation, mapping, and AI workloads.
CGRA Research
│
▼
Architecture Exploration
│
▼
CGLA Architecture
│
▼
FPGA RTL Prototype
│
▼
Mapping & Configuration
│
▼
AI Workloads
│
▼
Edge AI Demonstration
- What is a CGRA?
- CGRA vs FPGA, GPU and ASIC
- Processing Elements (PEs)
- Interconnect architectures
- Memory and data movement
- Configuration models
- Spatial computation
- Scheduling and mapping
- CGRA compilers
- MLIR and accelerator compilation
- AI/ML acceleration
- Low-power accelerator architectures
The CGLA architecture will serve as a simple experimental platform for exploring CGRA concepts.
Initial research will investigate:
- Processing Element design
- Linear PE arrays
- Inter-PE communication
- Configuration mechanisms
- Dataflow
- Local memory
- FPGA implementation
- Performance and resource utilization
- AI Workloads
After validating the basic architecture, the project will investigate mapping representative AI workloads onto the accelerator.
Potential targets include:
- Matrix and vector operations
- Neural-network kernels
- ONNX operators
- Small ML models
- Lightweight speech-processing workloads
- Moonshine and similar compact speech-recognition models
cgra-lab/
│
├── research/ # CGRA research and technical notes
├── architectures/ # Architecture definitions
│ └── cgla/
├── rtl/ # FPGA/RTL implementation
│ ├── pe/
│ ├── interconnect/
│ └── cgra/
├── simulation/ # Simulation and verification
├── compiler/ # Mapping and compilation experiments
├── models/ # AI models and model experiments
├── benchmarks/ # Performance and resource benchmarks
└── docs/ # Additional documentation
Phase 0 — Research
The project is currently in the research and architecture exploration phase.
The first objective is to build a solid understanding of CGRA architectures through primary research papers, technical publications, and open-source implementations before starting the FPGA prototype.
A curated list of academic papers, books, technical resources, and open-source projects will be maintained in:
research/references.md
Author: Cahit U. Ungan Focus: FPGA · Embedded AI · Computer Architecture · AI Accelerators