China can't buy Nvidia's RTX 4090. Now its optical chip is twice as fast

The Star

4 days ago

Chinese researchers have developed the first highly parallel optical computing integrated chip, named 'Meteor-1', setting a milestone for using light to perform an enormous number of operations at the same time, the scientists say.
The advance promises hardware acceleration for AI and data centres struggling with soaring computational demands.
The chip achieves a theoretical peak computing power of 2,560 TOPS (tera-operations per second) at 50GHz optical frequency – performance comparable to Nvidia's advanced GPUs – according to a report by Chinese publisher DeepTech last week.
Nvidia's latest GeForce RTX 5090 graphic card, for instance, peaks at 3,352 TOPS while its previous flagship RTX 4090 only reached 1,321 TOPS. In the past, optical chips remained mostly in laboratory settings, and could not come close to commercial flagship GPUs in real-life tasks.
Nvidia's 4090 and 5090 are effectively banned for sale to China because of US export controls on advanced semiconductors and AI chips that could aid Beijing in advancing its military capabilities.
As traditional electronic chips hit fundamental physical limits – from heat build-up, quantum effects and unsustainable power consumption – optical computing emerges as a critical future direction. Its inherent advantages, such as ultra-high speed, broad bandwidth, low power and minimal latency, position it to overcome these barriers.
Progress in optical computing has long focused on two key challenges: scaling up the matrix size and increasing optical frequency. Existing top models – exemplified by prototypes from TSMC and the California Institute of Technology – are pushing against both engineering and physical limits.
Consequently, a third way – expanding computational parallelism, or the ability of chips to multitask – has become the necessary path forward.
In a paper published on June 17 in eLight journal, Xie Peng and Han Xilin of the Shanghai Institute of Optics and Fine Mechanics (SIOM) and Hu Guangwei of Nanyang Technological University (NTU), Singapore, detailed a novel optical computing system that could hold more than 100 frequency channels in a single photonic chip.
'This achievement enables a 100-fold increase (and even beyond) in optical computility through ultra-high parallelism without scaling up the chip size, offering a novel technological pathway for future optical computers,' they said in the paper.
Han told DeepTech of the economic potential: 'This breakthrough promises to elevate optical computing to a cost-performance level competitive with electronic chips.'
The integrated Meteor-1 system features fully self-developed architecture, including a light source chip, optical interaction chip, optical computing chip and a modulation matrix driver board.
The light source chip uses an integrated micro-cavity optical frequency comb with an output spectrum exceeding 80 nanometres (nm), supporting more than 200 wavelengths. This chip-scale multi-wavelength source replaces hundreds of individual lasers, drastically reducing system size, power consumption and cost while boosting integration.
The core optical computing chip itself boasts a high transmission bandwidth over 40nm, enabling low-latency parallel processing. Complementing this, the team's custom-designed driver board features more than 256 channels for precise optical signal control and efficient processing.
Leveraging this system, the team set a world recording running more than 100 tasks at the same time on the system. Operating at 50GHz, the single chip delivers theoretical peak computing power of more than 2560 TOPS.
Leading researcher Xie Peng earned his PhD at Massachusetts Institute of Technology (MIT) in the United States and then conducted research at Oxford University and NTU Singapore before building on those research foundations to establish the optical chip and photonic computing team at SIOM in Shanghai last year.
Xie highlighted the advantage of the team structure at the Chinese Academy of Sciences to DeepTech: 'In our team, each key technical point has dedicated experts driving deep research. This modular-to-integrated approach enabled us to complete the full-chain innovation, from fundamental research to system integration, in a relatively short time frame.
'This also explains why challenges unresolved abroad for years saw breakthroughs quickly after my return.'
'Under our parallel optical computing scheme, key metrics like computational efficiency, power consumption and latency have the potential to surpass traditional electronic computing,' Xie said.
'We firmly believe optical computing, with the scalability challenge potentially addressed by our approach, can meet AI's ever-growing computational demands and unleash a wave of new applications.' – South China Morning Post