A cross-national team from China and Singapore has built what could be a glimpse of post-silicon computing. Their chip, recently announced as Meteor‑1, is a 100-wavelength optical processor that performs 2,500 multiply-accumulate operations per clock cycle using only light.
Powered by a soliton microcomb and a Mach–Zehnder Interferometer (MZI) mesh, this chip doesn’t use transistors for parallel processing. Instead, it encodes operations across 100 different wavelengths of light, each carrying a slice of the workload in parallel.
What’s Inside Meteor‑1?
Meteor‑1 is the real-world realization of the architecture described in the research paper “Parallel Optical Computing Capable of 100-Wavelength Multiplexing”, published in eLight in 2025 by teams from the Shanghai Institute of Optics and Fine Mechanics (SIOM) and Nanyang Technological University (NTU), Singapore.
At its core:
- A soliton microcomb generates over 100 coherent light channels.
- A 5×5 MZI mesh performs programmable matrix computations.
- A phase-correction framework ensures accuracy across all wavelengths.
Result: 2,500 parallel MAC operations per clock cycle, with over 90% spectral and matrix fidelity, and extremely low power consumption.
Introducing Meteor‑1: From Research to Hardware
While the original paper presents the experimental and architectural foundation, Chinese scientific and technology agencies have since named the chip Meteor‑1, marking it as the country’s first photonic AI accelerator chip.
Claimed specifications:
- Optical clock speed: 50 GHz
- Estimated peak performance: 2,560 TOPS (tera operations/second)
- Form factor: Integrated photonic chip (SOI platform)
- Power: Orders of magnitude lower than GPUs (primarily from pump lasers)
Meteor‑1 is seen as a direct response to increasing restrictions in semiconductor access and a strategic push toward AI hardware sovereignty.
How Meteor‑1 Compares to the World’s Most Powerful Chips
| Platform | Chip / Product | Compute Power | Power Draw | Year |
|---|---|---|---|---|
| Meteor‑1 (China) | Soliton Microcomb + 5×5 MZI × 100λ | 2,500 MACs/clock × 50 GHz = 2,560 TOPS | Low (optical) | 2025 |
| Nvidia | Blackwell B200 / GB200 | ~20 PFLOPS (FP4) | ~1–2 kW system | 2024 |
| GB300 Ultra | ~20 PFLOPS FP4 | ~700 W | 2025 | |
| RTX 5090 | ~4.2 TFLOPS Tensor FP16 | ~575 W | 2025 | |
| Lightmatter | Envise | 65.5 TOPS | 78 W + 1.6 W light | 2025 |
| Passage M1000 | 114 Tbps photonic interconnect | Low | 2025 | |
| Ayar Labs | TeraPHY | 8 Tbps optical I/O | Very low | 2025 |
| Intel Photonics | OCI Chiplet | 4 Tbps I/O | Moderate | 2024 |
Why Meteor‑1 Matters in the AI Hardware Race
While Nvidia’s Blackwell chips continue to dominate in raw AI performance, they come at a steep cost: enormous energy consumption, sophisticated cooling systems, and infrastructure complexity. These GPUs are built for brute-force processing, scaling through more transistors, more modules, and more power-hungry cores. It’s effective but not efficient at scale, especially as AI models surpass hundreds of billions of parameters.
In contrast, Lightmatter’s Envise photonic processor offers a promising middle ground. It can run full neural networks like BERT and ResNet using light instead of electrons. However, it scales in a more traditional way by adding more optical units (spatially), not spectrally. It’s production-ready, efficient, and impressive, but still tied to architectural constraints that require physical expansion.
Ayar Labs and Intel Photonics, on the other hand, are solving a different piece of the puzzle. Their focus is on optical interconnects, replacing copper links with ultra-fast photonic channels to reduce I/O bottlenecks in modern chiplet-based architectures. These are critical advancements, but they stop short of computing. They move data faster; they don’t compute it.
This is where Meteor‑1 introduces something fundamentally different. It brings a new scaling axis into play: light spectrum. Rather than adding more physical cores or chiplets, it unlocks massive parallelism using 100 unique wavelengths of coherent light, each wavelength performing independent matrix operations. It doesn’t need more space. It doesn’t need more power. It simply uses more “colors.”
Game-Changing Use Cases for Optical AI Compute
Meteor‑1 holds enormous promise for AI inference and acceleration, especially in transformer-based architectures. These models rely heavily on matrix multiplication, exactly the kind of operation Meteor‑1 performs optically, in parallel, and at low energy cost. Optical multiply-accumulate (MAC) units like those in Meteor‑1 could potentially replace the most power-hungry stages in AI pipelines.
The architecture also opens the door to low-latency, low-power computing at the edge. Because optical systems don’t require clocked logic and consume little heat, they can power passive or near-passive AI processors. In the future, chips like Meteor‑1 could enable real-time AI on satellites, drones, autonomous vehicles, or even smartphones, places where traditional GPUs are either impractical or unsustainable.
Beyond applications, Meteor‑1 carries geopolitical weight. This chip isn’t just a technical breakthrough; it signals China’s strategic intent to build sovereign hardware capabilities. In a time of rising restrictions on semiconductor exports, Meteor‑1 represents a bid for technological independence and leadership in the most critical layer of the AI stack: compute.
What Meteor‑1 Still Lacks and Where It’s Headed
As revolutionary as it is, Meteor‑1 is still an early prototype with limitations. First, it performs only linear matrix operations. There’s no on-chip implementation of nonlinear activation functions (like ReLU or GELU), which are essential for full neural network execution. This means it must be paired with external electronic systems to complete a full model pipeline.
Second, phase programming is voltage-controlled, requiring external drivers to tune the MZI mesh. While functional in lab conditions, this approach needs to evolve into a compact, software-controllable format for broader adoption.
Third, integration with optical memory or high-speed interconnects remains under development. Today’s architecture processes data with light, but it still relies on electronic subsystems to move and store information.
That said, the fundamental physics have been validated. The system has already proven high fidelity and spectral consistency across 100 channels. Future generations could expand to 200+ wavelengths, integrate with nonlinear photonics, and become the backbone of hybrid electro-photonic neural accelerators.
Discover more from Semiconductors Insight
Subscribe to get the latest posts sent to your email.
1 thought on “Meteor‑1: China’s Latest Optical AI Chip That Competes with NVidia Through Colors”