英伟达引爆CPO新战场

Core Viewpoint - NVIDIA has officially launched two CPO (Co-Packaged Optics) switch products, Quantum-X Photonics for InfiniBand and Spectrum-X Photonics for Ethernet, with the InfiniBand CPO expected to debut in the second half of 2025 and the Ethernet CPO in the second half of 2026 [1][2]. Group 1: CPO Technology and Market Position - CPO will be an optional configuration, and NVIDIA will continue to offer traditional switch systems with pluggable modules [2]. - The primary driver for NVIDIA's investment in CPO technology is power optimization, with a significant reduction in power consumption from 30W to 9W for a 1.6T port, achieving a 70% decrease [2]. - NVIDIA's CPO solution utilizes new micro-ring modulators (MRM) for enhanced energy efficiency, contrasting with Broadcom's CPO solution that achieved a 50% power reduction using traditional Mach-Zehnder modulators (MZM) [2]. Group 2: Ecosystem and Partnerships - The technology involves a multi-component ecosystem, integrating electronic and photonic chips through 3D stacking, with TSMC's compact optical engine technology playing a key role [3]. - NVIDIA's CPO partners include major industry players such as Browave, Coherent, Corning, Fabrinet, Foxconn, Lumentum, Senko, SPIL, Sumitomo, Tianfu Communication, and TSMC [3]. Group 3: Strategic Implications and Future Developments - LightCounting notes that NVIDIA's first CPO product is an InfiniBand switch, which has been overshadowed by Ethernet in NVIDIA's AI strategy, indicating that the initial deployment will primarily serve NVIDIA's own clusters [4]. - The Spectrum-X platform aims to elevate Ethernet performance to match that of InfiniBand, with potential for significant GPU interconnectivity in future architectures [4]. - NVIDIA's entry into the CPO market is expected to invigorate the ecosystem, with both NVIDIA and Broadcom projected to release single-channel 200G CPO switches by 2027, leading to a more mature industry landscape [4]. Group 4: Technical Challenges and Innovations - Scale-out is seen as a low-risk entry point for CPO, while scale-up is critical for success, particularly for the mixed expert (MoE) model requiring rapid response times across GPUs [5]. - NVIDIA has previously announced fiber-based NVLink plans, with at least one cluster built internally, but large-scale deployment has been hindered by high power consumption of timing modules [6]. - The NVLink CPO is scheduled for 2028, allowing NVIDIA to validate technology feasibility over two product cycles, significantly reducing future integration risks [6].