In a potential shift in how future AI systems are powered, researchers from Tsinghua University in Beijing have unveiled a ground-breaking all-analogue photoelectronic chip that significantly outperforms conventional digital processors in both speed and energy efficiency.
Known as ACCEL (All-Analogue Chip Combining Electronic and Light Computing), the processor represents a radical departure from the binary 0s and 1s of standard silicon, utilising the physical properties of light to perform computations at unprecedented speeds.
The results, published in the journal Nature, suggest that ACCEL can perform vision-related AI tasks—such as image recognition and autonomous driving calculations at a processing rate nearly 3,000 times faster than NVIDIA’s widely used A100 GPU, as reported by SCMP.
Even more striking is its energy profile; the team reportedly claims the chip operates with an efficiency of 74.8 peta-operations per second per watt, meaning the energy required to run a standard digital chip for one hour could power ACCEL for over five centuries.
Real-world performance gains and implications for AI hardware
In benchmark tests using data sets comparable in size to those of large online platforms such as Netflix and Yahoo, the analogue chip demonstrated dramatic improvements over digital counterparts, according to reports.
In one application, it completed recommendation system training tasks far faster than a typical digital processor while consuming a fraction of the power. In image-compression tests, results showed almost the same visual quality as traditional digital computations but with significantly reduced energy and storage requirements.
Energy efficiency has become a central challenge in the AI sector, where large neural networks and data centres consume vast amounts of power. Modern AI training clusters powered by advanced GPUs can demand hundreds of watts per device, creating both cost and sustainability concerns for developers and cloud operators alike. An analogue architecture that delivers high performance with dramatically lower energy consumption could help reshape this landscape, offering a route to greener, more scalable AI implementations.
The breakthrough builds on earlier analogue and hybrid innovations by Chinese research teams, including work showing analogue computing systems could, in theory, outperform current GPUs by orders of magnitude in specific problem domains. By overcoming historical limitations in precision and practical real-world application, the latest design shows that analogue computing may be moving from the lab into feasible AI hardware solutions.
A strategic “lane-change” for China
The development comes at a critical time for China’s semiconductor industry, which faces stringent US export controls on high-end GPUs. Interestingly, ACCEL does not require the most advanced sub-5nm fabrication processes, which are currently restricted. Instead, it was manufactured using older, more accessible transistor technology, proving that architectural innovation can sometimes bypass the need for cutting-edge lithography.
While the chip is currently specialised for vision tasks and lacks the general-purpose versatility of a standard CPU or GPU, its performance in specific AI benchmarks, reaching over 90% accuracy in video recognition, marks it as a potent specialised accelerator for the next generation of smart devices.
