Not exactly a DeepSeek moment for AI accelerators - but this Chinese optical chip may well be 100x faster than Nvidia's A100 on some tasks
Chinese photonic AI chips demonstrate extreme speed and energy efficiency in narrowly defined generative workloads, exceeding conventional GPU performance in lab tests.
- Chinese photonic chips reportedly outperform conventional GPUs in narrow, specialized generative AI tasks
- ACCEL combines photonic and analog electronic components to achieve high computational throughput
- LightGen uses over two million photonic neurons for all-optical generative AI processing
Chinese research institutions have described new photonic AI chips that allegedly outperform conventional GPUs by a very large margin under specific conditions.
The institutions and researchers claim these chips show dramatic upgrades in speed and energy efficiency when executing narrowly defined generative workloads.
They reported China’s light-based AI chips offer 100x faster speed than Nvidia GPUs at some tasks, particularly in areas such as image synthesis, video generation, and vision-related inference.
Claims of extreme speed from optical computing research
These claims are framed around laboratory evaluations rather than commercial deployment scenarios, but the performance gap is closely tied to fundamental architectural differences.
Nvidia GPUs, including widely used accelerators such as the A100, rely on electronic circuits where electrons move through transistors to execute programmable instructions.
This approach allows flexibility across many workloads but results in high power consumption, significant heat output, and dependence on advanced manufacturing nodes.
The Chinese photonic chips instead rely on light-based signal processing, where photons replace electrons as the medium for computation, enabling massive parallelism through optical interference rather than sequential digital execution.
Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed!
One of the reported chips, ACCEL, was developed at Tsinghua University as a hybrid system combining photonic components with analog electronic circuitry.
It reportedly operates using older semiconductor manufacturing processes while achieving extremely high theoretical throughput figures measured in petaflops.
These calculations are limited to predefined analog operations rather than general-purpose code execution.
ACCEL is therefore designed for tasks such as image recognition and vision processing.
These workloads rely on fixed mathematical transformations and tightly controlled memory access patterns.
A second system, LightGen, was developed through collaboration between Shanghai Jiao Tong University and Tsinghua University.