- 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.
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.
LightGen is described as an all-optical computing chip containing more than two million photonic neurons.
Research papers claim it can perform generative tasks such as image generation, denoising, three-dimensional reconstruction, and style transfer.
Experimental results reportedly show performance gains exceeding two orders of magnitude when compared with leading electronic accelerators.
These measurements are based on time and energy consumption under constrained conditions.
These systems are not described as replacements for GPUs in general computing, training large models, or executing arbitrary software.
They function as specialized analog machines designed for narrow categories of computation.
The reported claims suggest that optical computing can deliver exceptional gains when workloads are carefully shaped to fit the hardware.
The gap between lab demonstrations and usable AI tools remains large, with task-specific versus general-purpose capability central to evaluating these claims.
Via Interesting Engineering | China Daily | EurekAlert
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
https://cdn.mos.cms.futurecdn.net/THrQgcJizbtksGzNr9f3qb-1920-80.jpg
Source link
