Sustainable Hybrid (CPU + PIM) Computing Architecture for Next-Gen AI Workloads
"Moving the processor to the data, instead of moving data to the processor."
View Demo • Documentation • Report Bug
Green-Tensor-Core is a simulation framework for a Hybrid Computing Architecture designed to tackle one of the biggest challenges in modern computing: Data Movement Energy Costs.
In traditional Von Neumann architectures, up to 62.7% of total system energy is wasted on moving data between Memory (DRAM) and the Processor (CPU/GPU).
This project proposes a sustainable solution by integrating Processing-in-Memory (PIM) accelerators with a host CPU.
By offloading heavy vector-matrix operations to the memory module and keeping control logic on the CPU, Green-Tensor-Core aims to drastically reduce the carbon footprint of:
- 5G/6G Networks
- Green Data Centers
- Autonomous Systems
- Edge AI & IoT Devices
The system follows a heterogeneous computing model with three main layers:
📌 (Place your architecture diagram in
docs/architecture_diagram.pngand reference it here)
- Role: Manages OS, I/O requests, and light serial processing
- Logic: Acts as the "brain" and decides which tasks should be offloaded
- Role: Performs high-intensity parallel computations (e.g., Deep Learning Inference, FFT) directly inside memory
- Benefit: Near-zero data movement cost for large datasets
Core Innovation:
A runtime scheduler that analyzes incoming tasks:
- If task is Data-Intensive (e.g., Matrix Multiplication) → Offload to PIM
- If task is Control-Intensive (e.g., Branching Logic) → Execute on CPU
| Domain | Problem | Green-Tensor-Core Solution |
|---|---|---|
| 📡 5G & 6G Networks | Base stations consume massive power for signal processing | Reduces energy per bit via in-memory processing |
| ☁️ Green Data Centers | AI training (LLMs) causes high heat & carbon emissions | Lowers cooling cost and TDP |
| 🛸 Autonomous Systems | Limited battery for AI workloads | Extends flight time / driving range |
| 🔒 Edge AI & IoT | Cloud offloading is slow and risky | Enables secure, low-latency on-device AI |
Benchmarked against CPU-only architectures using synthetic workloads (ResNet-50-like matrix operations):
- ⚡ Energy Savings: ~42% reduction in total energy consumption
- ⏱️ Latency: 1.8× speedup for large batch workloads
- 📉 Bus Utilization: 60% reduction in memory bus traffic
- Python 3.8+
pip
git clone https://github.com/bugradba/Green-Tensor-Core.git
cd Green-Tensor-Core
pip install -r requirements.txt
python src/main.py --mode hybrid --workload large_matrix
Green-Tensor-Core/
├── PIM SIMULATOR/
│ ├── Q_Learning/ # Q-Learning based scheduling utilities
│ │ ├── (Q_Learning içerikleri) # Q-learning scripts & helpers
│ │
│ ├── adaptive_scheduler.py # Adaptive scheduler module
│ ├── baseline_models.py # Baseline (CPU-only) models
│ ├── hybrid_scheduler.py # Hybrid CPU + PIM scheduler logic
│ ├── pim_simulator.py # Core processing-in-memory simulation engine
│ ├── adaptive_precision.py # Precision/energy trade-off experiments
│ ├── test_simulator.py # Simulation tests & benchmarks
│ └── visualize_scheduler.py # Visualization tools for scheduler behavior
├── README.md # Project overview and docs
└── requirements.txt # Python dependencies
The final hardware and software architecture of the project is built upon a heterogeneous computing framework. The system is controlled by a Q-Learning-based Smart Hybrid Scheduler that manages workload distribution between the CPU and PIM (Processing-in-Memory) units.
The architectural diagram below illustrates the details of this integration:
To overcome the Memory Wall bottleneck, the scheduler dynamically analyzes tasks (profiling) and routes them based on the following logic:
IF (\text{data_intensive}) \rightarrow \text{PIM (Processing-in-Memory)} ELSE \rightarrow \text{CPU (Host Processor)}
###Contact & Acknowledgements
-
**Developer: Muhammed Buğra Demirbaş
-
**Context: Developed for Tomorrow's Technology Leaders (Sustainability Track).
-
**LinkedIn: https://www.linkedin.com/in/m-bugra-demirbas/
-
**Email: mbugrademirbas@gmail.com