China Unveils Photonic AI Chips with Massive Performance Claims
China has unveiled another generation of AI microchips that claim great performance advantages over competing hardware in certain scenarios. These chips employ light (photons) instead of electrons as in traditional GPUs from NVIDIA amongst others.
An Alternative Approach to AI Processing
Conventional GPUs are general-purpose processors, but their electron-based transistor architecture generates a lot of heat and therefore powers a lot. Photonic chips are built on a different architecture aimed at performing quite specific tasks of AI such as computer vision and image generation, with a plus point of being more energy efficient.
Two Photonic Chip Development Initiatives
Information has been disclosed with respect to two separate chips:
ACCEL A Hybrid Photonic-Electronic Chip
- Architecture: Using a hybrid combination of photonic components and analog electronics
- Performance: Claimed by its creators to reach up to 4.6 petaflops
- Efficiency: Operates with extremely low energy consumption
- Functionality: Chip does not run software code; instead, it runs a few predefined analog operations
- Manufacturing: Can be manufactured with older SMIC manufacturing technology
LightGen An All-Photonic Chip
- Architecture: A fully photonic chip with more than 2 million photonic "neurons."
- Application: Support image generation and image processing tasks.
- Performance: It is reported to be operating at least 100 times faster than conventional GPUs on its assigned tasks.
While these photonic solutions will not be direct replacements for GPU versatility, they clearly signal some technological leaps toward greater speed and efficiency for an increasing number of specific AI tasks.
