8 Phoenix is a hardware-software co-design project aiming to solve the AI "memory wall" using light. By integrating silicon photonics directly onto a 7nm compute tray, we enable 8 chips to act as a single, unified 1TB processor. Open-sourced to provide a high-performance alternative to proprietary AI hardware.
Current AI hardware is locked behind proprietary ecosystems and "leading-edge" price tags. 8 Phoenix changes the game by:
- 7nm Maturity: Utilizing the high-yield, lower-cost 7nm node for sustainable manufacturing.
- Optical Fabric: Replacing power-hungry copper with light-speed Silicon Photonics for chip-to-chip communication.
- Open Architecture: Providing a bilingual (English/Chinese) open-source repository for the global community to inspect, build, and scale.
| Feature | Specification |
|---|---|
| Process Node | 7nm (FinFET) |
| Compute Units | 8-Chip Synchronous Tray |
| Peak Performance | 2.4 PetaFLOPS (FP16/BF16) |
| Unified RAM | 1,024 GB (1 TB) via HBM3 |
| Interconnect | Phoenix-Link Optical Mesh (3.2 TB/s per chip) |
| TDP | 300W per Unit |
docs/: Bilingual technical specifications and ISA manuals.hw/: RTL designs, floorplans, and memory controller logic.sw/: Compiler (Phoenix-C), functional simulators, and driver stacks.examples/: "Hello World" matrix math and basic AI model implementation.
Read the Instruction Set Architecture (ISA) to understand how 8 Phoenix processes matrix operations deterministically.
Test the execution logic of the Phoenix-Matrix instructions using our Python-based simulator:
python sw/simulator/functional_model.py