NVIDIA (NASDAQ:NVDA) used a recent webinar on “Co-package Silicon Photonic Switch for Gigawatt AI Factories” to outline how it is designing networking and optical interconnects to support large-scale AI “factories,” with Senior Vice President of Networking Gilad Shainer detailing the company’s approach to building AI supercomputers and the role co-packaged optics (CPO) plays in power efficiency and reliability.
Data centers as “the computer” and four interconnect layers
Shainer framed modern AI infrastructure around the idea that “the data center is the computer,” arguing that AI workloads depend on many computing elements operating together and that the network increasingly defines the capability of the overall system.
- Scale-up: NVLink connects NVIDIA H100 GPUs to form what he called a rack-scale GPU, with tens of GPUs today and potentially hundreds in the future operating as a single unit.
- Scale-out: Spectrum-X Ethernet is positioned as the end-to-end fabric connecting racks to enable distributed workloads across hundreds of thousands of GPUs.
- Context memory storage: A storage tier within an AI pod, leveraging BlueField DPUs “data and storage processors,” aimed at inferencing storage requirements.
- Scale across: A Spectrum-X-based approach intended to connect multiple data centers into a single computing engine when single-site scaling is limited by power or real estate, targeting “giga-scale” AI factories.
Spectrum-X Ethernet: focus on low jitter and distributed AI
In discussing Spectrum-X, Shainer said Ethernet in the market has historically been optimized for enterprise data centers, hyperscale single-server workloads, or service provider environments—none of which, he argued, were designed specifically for AI distributed computing at very large scale. He said Spectrum-X runs standard Ethernet protocols, but is engineered as an end-to-end system to reduce jitter and improve synchronization across GPUs.
He emphasized two key architectural elements:
- SuperNICs that control injection rate to help avoid hotspots (which he said are a source of jitter) and place data in the correct GPU memory location.
- Switch behavior that uses fine-grain adaptive routing, selecting paths on a packet-by-packet basis based on local and remote conditions.
Shainer cited performance claims tied to jitter reduction. For inference workloads, he said Spectrum-X improves “expert dispatch performance” (an all-to-all operation sensitive to jitter) by 3x. For training, he said the platform provides 1.4x performance along with predictable step times, describing predictability as a key requirement for AI workloads.
Why co-packaged optics: power and resiliency
Shainer then turned to the power impact of optical connectivity in large GPU clusters. As bandwidth doubles generation to generation, he said power consumption of the optical network rises and can approach roughly 10% of computing resources. Reducing optical power, he argued, can translate into higher compute capacity within power-constrained data centers.
He described co-packaged optics as moving the optical engine—typically housed in an external transceiver—into the same package as the switch ASIC. By shortening the electrical path and reducing transitions, NVIDIA expects lower power consumption and improved signal quality. Shainer referenced pluggable transceivers in the 20–25 watt range and said co-packaged optics can deliver up to a 5x power savings for the scale-out infrastructure.
Beyond power, he said co-packaged optics reduces components and the number of lasers required, improves data center resiliency, and increases “time to first interrupt.” NVIDIA has built CPO-capable switches for both Spectrum-X Ethernet and Quantum-X InfiniBand, he said, and worked with an ecosystem of partners on packaging, fiber attach methods, and liquid-cooled designs.
Switch specifications and stated technology metrics
Shainer provided a number of switch configurations and technology assertions tied to the photonics roadmap:
- Spectrum-X Ethernet Photonics: 200G SerDes co-packaged optics; Shainer said it doubles bandwidth versus the prior generation. He also claimed 64x higher signal integrity, 13x higher laser reliability, and significant component reduction by removing pluggable transceivers.
- Quantum-X InfiniBand Photonics: a 115-terabit switch with 144 ports of 800G, described as fully liquid-cooled for power efficiency.
- Spectrum-X switch configurations: a 102-terabit switch with 120 ports of 800G or 512 ports of 200G; and a larger 409 terabit-per-second switch with 512 ports of 800G or 2,000 ports of 200G.
Deployment timing, reliability questions, and hyperscaler adoption considerations
During Q&A, Shainer said NVIDIA expects “co-package optics deployments” to begin this year. He said the company had announced three partners—CoreWeave, Lambda, and Texas Advanced Computing Center—that will deploy Quantum-2 InfiniBand co-packaged optics in the first part of the year. He added that Spectrum-X Ethernet co-packaged optics would begin shipping in the second part of the year.
On reliability, Shainer pointed to human handling of pluggable optics—cleaning, insertion, and accidental contact—as a contributor to transceiver replacement needs. He said integrating optical engines inside the switch package and validating the system as a whole reduces exposure to dust and handling, supporting higher resiliency.
Asked about collaboration with TSMC, Shainer highlighted co-packaging processes designed for reliability and testability. He also described changes to optical engine design, saying earlier CPO attempts used larger MZM-based engines, while NVIDIA’s approach uses a micro-array modulator to support large radix switches. He added that NVIDIA developed fiber alignment approaches and high-power lasers intended to reduce the number of lasers required.
On flexibility versus pluggable optics, Shainer said CPO requires selecting a specific technology, but argued NVIDIA’s implementation covers data center distances and can connect remote buildings on a campus, reducing the need for multiple transceiver types within the data center. For very long-distance data center-to-data center connections, he said transceivers would still be used.
Finally, responding to concerns about the “pay-as-you-go” nature of pluggable optics, Shainer said AI supercomputers are designed for high utilization and optimized topologies, suggesting that in such builds customers would typically populate the infrastructure fully, and that CPO can reduce both capital and operating costs while improving resiliency.
About NVIDIA (NASDAQ:NVDA)
NVIDIA Corporation, founded in 1993 and headquartered in Santa Clara, California, is a global technology company that designs and develops graphics processing units (GPUs) and system-on-chip (SoC) technologies. Co-founded by Jensen Huang, who serves as president and chief executive officer, along with Chris Malachowsky and Curtis Priem, NVIDIA has grown from a graphics-focused chipmaker into a broad provider of accelerated computing hardware and software for multiple industries.
The company’s product portfolio spans discrete GPUs for gaming and professional visualization (marketed under the GeForce and NVIDIA RTX lines), high-performance data center accelerators used for AI training and inference (including widely adopted platforms such as the A100 and H100 series), and Tegra SoCs for automotive and edge applications.
