China debuts its Meteor-1 chip to compete with Nvidia amid US export restrictions

Scientists in China have rolled out “Meteor-1,” reportedly the nation’s inaugural highly parallel optical computing chip, marking an important leap in using light-based hardware to handle huge parallel workloads.

The initiative comes as the country is weaning itself off US-made technology amid export restrictions of advanced AI tech. The local media in China reported that this is expected to result in hardware acceleration for AI and data centers currently struggling with rising computational demand.

China’s new chip competes with Nvidia’s top GPU

Designed by teams at the Shanghai Institute of Optics and Fine Mechanics (SIOM) and Nanyang Technological University, the device reportedly boasts a theoretical peak performance of 2,560 TOPS (tera-operations per second) at a 50 GHz optical clock.

That places it in the same ballpark as Nvidia’s top GPUs, offering a home-grown solution to accelerate AI and data center tasks amid ballooning computational demands and US chip restrictions.

Historically, optical computing efforts have centered on enlarging the dimensions of the internal matrix that performs calculations.

Bigger matrices theoretically allow more data to be processed in parallel, but in practice, they run headlong into engineering constraints, complex chip layouts, the need for extreme precision in waveguide alignment and prohibitive fabrication costs.

According to the South China Morning Post (SCMP), efforts by TSMC and academic groups such as Caltech have shown promise in laboratory settings, but those prototypes have struggled to translate into commercial-grade solutions.

Another strategy has been to push the oscillation rate of the light itself. Higher optical frequencies can deliver faster computation, but they also amplify signal losses, exacerbate thermal noise and raise the bar on component tolerances.

Until now, no team has managed to pair both large matrix scales and ultra-fast optics in one system without running up against severe trade-off ceilings that erode real-world performance.

Meteor-1 handles complex tasks and is the answer to US sanctions

Meteor-1 charts a different course; multiplying the number of simultaneous tasks rather than enlarging individual components. The June 17 paper in eLight by Xie Peng, Han Xilin and Hu Guangwei outlines how the chip incorporates over 100 distinct frequency channels within one photonic platform.

This high-order parallelism enables a hundredfold, or greater, increase in “optical computility” without expanding the chip’s footprint, delivering a practical path for next-generation light-powered processors.

With US export curbs effectively banning Nvidia’s RTX 4090 (1,321 TOPS) and RTX 5090 (3,352 TOPS) from China, the Meteor-1 effort arrives at a critical juncture.

Conventional electronic semiconductors are bumping into fundamental limits, heat dissipation, quantum tunnelling and unsustainable power draws. Optical chips sidestep many of these barriers, offering ultra-high speed, broad bandwidth, reduced energy consumption and minimal latency.

Meteor-1’s architecture is entirely home-designed. Its on-chip light source uses a micro-cavity optical frequency comb that covers more than 80nm of spectrum, spanning upwards of 200 wavelengths. This innovation effectively replaces hundreds of discrete lasers, slashing system complexity, power demands and costs.

The core computing die itself offers a transmission bandwidth above 40nm, facilitating low-latency, massively parallel operations. Coupled with a bespoke driver board featuring over 256 channels for precise signal modulation, the system executed more than 100 simultaneous tasks in benchmark tests, all at a 50 GHz clock, yielding 2,560 TOPS.

Han Xilin tells DeepTech that Meteor-1’s cost-performance metrics could soon rival electronic GPUs. Lead researcher Xie Peng, a Massachusetts Institute of Technology PhD who went on to research at Oxford and NTU, attributes the rapid progress to SIOM’s modular team structure under the Chinese Academy of Sciences.

“Each critical subsystem had its own dedicated expert, allowing us to integrate full-chain innovation from foundational research through system assembly in a remarkably short span.”

~ Han Xilin.

Looking ahead, the group believes their parallel-first design could outpace electronic chips on efficiency, power draw and latency, meeting AI’s insatiable compute appetite while spawning novel applications in real-time data analysis, autonomous systems and scientific modeling.

Cryptopolitan Academy: Tired of market swings? Learn how DeFi can help you build steady passive income. Register Now