NVIDIA launches new silicon photonics platform: supports millions of GPU clusters, based on TSMC COUPE technology

NVIDIA launches groundbreaking silicon photonics technology, creating a million GPU scale cluster. At GTC 2025, NVIDIA officially released the Spectrum-X Photonics and Quantum-X Photonics network switch platforms for ultra-large-scale data centers. These new network switch platforms, utilizing silicon photonics technology, achieve data transmission speeds of up to 1.6Tb/s per port, totaling 400Tb/s, enabling seamless collaboration among millions of GPUs. NVIDIA stated that compared to traditional network solutions, these new switches offer higher bandwidth, lower power loss, and superior reliability. The Quantum-X InfiniBand switch is expected to be released later in 2025, while the Spectrum-X Photonics Ethernet switch will debut in 2026. Quantum-X performance doubles, AI computing scalability increases fivefold The Spectrum-X Photonics Ethernet and Quantum-X Photonics InfiniBand platforms achieve speeds of up to 1.6Tb/s per port (twice the maximum speed of current top copper Ethernet solutions), with total bandwidth reaching 400Tb/s through various port configurations. NVIDIA's Spectrum-X Photonics switches offer multiple configurations, providing 128 ports at 800Gb/s or 512 ports at 200Gb/s, with a total bandwidth of 100Tb/s; higher capacity models offer 512 ports at 800Gb/s or 2048 ports at 200Gb/s, achieving 400Tb/s throughput. The Quantum-X Photonics series features 144 ports at 800Gb/s InfiniBand, utilizing 200Gb/s SerDes for efficient data transmission. Compared to the previous generation of network solutions, Quantum-X doubles performance and increases AI computing scalability fivefold, making it suitable for high-intensity workloads and building larger AI clusters. Additionally, the Quantum-X InfiniBand switch employs a liquid cooling system, ensuring that the onboard silicon photonics chips operate at peak efficiency without overheating. Therefore, these new network platforms promise 3.5 times better energy efficiency, 10 times higher network reliability, and 63 times stronger signal integrity, reducing power consumption and enhancing long-term performance. Utilizing TSMC's COUPE technology, further follow-ups may continue Furthermore, it was reported that NVIDIA's new silicon photonics platform uses TSMC's Compact Universal Photonic Engine (COUPE) silicon photonics platform, which combines 65nm electronic integrated circuits (EIC) with photonic integrated circuits (PIC), utilizing the company's SoIC-X packaging technology Analysis indicates that although NVIDIA has not disclosed its future plans for network equipment, TSMC's COUPE technology roadmap is very promising, and NVIDIA may follow suit. NVIDIA is also collaborating with companies such as Coherent, Corning, Foxconn, Lumentum, and Senko to establish its own silicon photonics ecosystem, with a stable supply chain, enabling NVIDIA and its partners to build AI clusters and data centers that were previously unattainable with proprietary hardware. It is reported that TSMC's next-generation silicon photonics will integrate COUPE technology in CoWoS packaging, combining optics directly with switches. This will allow optical connection speeds to reach 6.4Tb/s. The third version aims to increase the speed to 12.8Tb/s, directly integrated into processor packaging. Media analysis states that while these advancements promise significant improvements, integrating silicon photonics technology at such a large scale is complex, and widespread adoption depends on whether organizations are willing to upgrade their existing network infrastructure