NVIDIA's silicon photonics switches. [Image: NVIDIA]
The US-based artificial intelligence (AI) computing multinational NVIDIA has announced its plan to leverage silicon photonics and co-packaged optics (CPO) to connect millions of GPUs in so-called AI factories. At the company’s GTC 2025 event in San Jose, CA, USA, CEO Jensen Huang introduced two new silicon photonics network switch platforms, Spectrum-X and Quantum-X, that he said would enable this leap while also reducing power consumption. “NVIDIA has achieved the fusion of electronic circuits and optical communications at massive scale,” noted the company.
Bigger and more efficient
The rapid growth of AI technology has put unprecedented demands on networks and data centers. NVIDIA says that its new photonic switches are “the world’s most advanced networking solution” to this problem. “AI factories are a new class of data centers with extreme scale, and networking infrastructure must be reinvented to keep pace,” said Huang. “By integrating silicon photonics directly into switches, NVIDIA is shattering the old limitations of hyperscale and enterprise networks and opening the gate to million-GPU AI factories.”
According to NVIDIA, an AI factory can use up to 2.4 million optical transceivers and up to 24 MW of power, which can be upwards of 10% of the overall data center power consumption. Huang claimed that NVIDIA’s silicon photonics technology could save data centers “tens of megawatts” of power consumption.
To meet the lofty million-GPU goal while using less energy, the platforms use 1.6 Tb/s port switches, and they leverage CPO for short-range connections to improve overall electrical efficiency. The company says that the switches “deliver 3.5× more power efficiency, 63× greater signal integrity, 10× better network resiliency at scale and 1.3× faster deployment compared with traditional methods.” The CPO is based on a new micro ring modulator technology that NVIDIA claims offers an extra power savings.
A look at the specs
The Quantum-X Photonics switches provide 144 ports of 800 Gb/s InfiniBand and use a liquid-cooled design to cool the onboard silicon photonics. The package combines 18 silicon photonics engines, enabling 324 optical connections and 288 data links from 36 laser inputs. Connected to the package are six detachable optical sub-assemblies, each comprising three silicon photonics engines to provide an aggregate data throughput of 4.8 Tb/s.
The Spectrum-X ethernet networking platform includes multiple photonics configurations, including 128 ports of 800 Gb/s or 512 ports of 200 Gb/s, delivering 100 Tb/s total bandwidth, as well as 512 ports of 800 Gb/s or 2048 ports of 200 Gb/s, for a total throughput of 400 Tb/s. The Quantum-X switches are expected to be available later this year, while the Spectrum-X switches will come in 2026.
Working together
NVIDIA has teamed up with a number of partners in its silicon photonics efforts. Photonics firm Lumentum is providing the lasers for Spectrum-X, while Coherent is collaborating on CPO. Both companies were already key suppliers of optical transceivers to NVIDIA. The company’s “silicon photonics econsystem” also includes Browave, Corning, Fabrinet, Foxconn, Senko, SPIL, Sumitomo Electric, TFC and TSMC.
“A new wave of AI factories requires efficiency and minimal maintenance to achieve the scale required for next-generation workloads,” said C. C. Wei, chairman and CEO of TSMC. “TSMC’s silicon photonics solution combines our strengths in both cutting-edge chip manufacturing and TSMC-SoIC 3D chip stacking to help NVIDIA unlock an AI factory’s ability to scale to a million GPUs and beyond, pushing the boundaries of AI.”