NVIDIA Spectrum-X Photonics
Last reviewed
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,913 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,913 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA Spectrum-X Photonics is a co-packaged optics (CPO) Ethernet switch platform from Nvidia, announced at the company's GPU Technology Conference (GTC) on March 18, 2025. It integrates silicon-photonics optical engines directly into the switch package alongside the switch ASIC, replacing the bulk of the pluggable optical transceivers used in conventional networking. By eliminating those discrete transceivers and the electrical retiming and signal-conditioning circuitry that feed them, Spectrum-X Photonics is designed to cut interconnect power, raise reliability, and let Ethernet fabrics scale toward AI data centers containing hundreds of thousands to millions of GPUs. It is the Ethernet member of NVIDIA's silicon-photonics switch family; its sibling, Quantum-X Photonics, applies the same approach to InfiniBand. [1][2]
The platform extends NVIDIA's existing Spectrum-X Ethernet networking line, which pairs Spectrum switch silicon with NVIDIA SuperNIC and adapter technology to provide a lossless, AI-optimized Ethernet fabric. Spectrum-X Photonics keeps that fabric architecture but changes the physical layer: instead of connecting the switch ASIC to faceplate cages full of pluggable modules, the optics are co-packaged with the silicon so that fiber attaches much closer to the chip. [2][3]
In a traditional data center switch, the switch ASIC drives high-speed electrical signals (SerDes lanes) across the printed circuit board to the front panel, where pluggable transceivers (such as OSFP or QSFP modules) convert the electrical signal to light for transmission over optical fiber. Each pluggable module contains its own laser, modulator, and a digital signal processor (DSP) that retimes and cleans up the electrical signal after its journey across the board. As lane rates climb to 200 Gb/s and beyond, that electrical path becomes increasingly lossy and power-hungry, and the large population of pluggable modules and their lasers becomes a significant source of both power draw and field failures. [1][4]
Co-packaged optics moves the electrical-to-optical conversion from the faceplate to a position immediately adjacent to the switch ASIC, inside the same package. The optical engines sit beside the switch silicon on a shared substrate, so the high-speed electrical signals travel only a very short distance before being converted to light, and fiber is brought directly to the package. Because the link no longer has to survive a long, lossy board trace and a connector, the per-lane DSP retimers used in pluggable optics can be removed, which lowers power and latency and reduces the number of components that can fail. NVIDIA frames this as a prerequisite for "AI factories" at the million-GPU scale, where the aggregate power consumed by pluggable optics and the failure rate of millions of modules would otherwise become limiting factors. [1][4][5]
NVIDIA introduced Spectrum-X Photonics alongside Quantum-X Photonics during CEO Jensen Huang's GTC 2025 keynote, positioning them as the world's first 1.6 Tb/s-class co-packaged silicon-photonics networking switches built for AI infrastructure. NVIDIA stated that the platforms integrate optics innovations that, compared with conventional pluggable-transceiver networking, deliver 3.5 times better power efficiency, 63 times greater signal integrity, 10 times better network resiliency at scale, and 1.3 times faster deployment, while using 4 times fewer lasers. [1]
NVIDIA described the optical engines as built on a 3D-stacked silicon-photonics process developed with TSMC, combining TSMC's advanced logic manufacturing with its SoIC 3D chip-stacking and advanced packaging to fuse the photonic and electronic layers. Jensen Huang summarized the goal by saying NVIDIA was "integrating silicon photonics directly into switches" to break "the old limitations of hyperscale and enterprise networks" and open "the gateway to AI factories with millions of GPUs." [1]
Spectrum-X Photonics is built around NVIDIA's Spectrum-6 switch ASIC. Rather than feeding the ASIC's SerDes out to faceplate transceiver cages, the design surrounds the switch silicon with co-packaged silicon-photonics optical engines and brings optical fiber directly to the package through detachable fiber connectors. NVIDIA has described the larger Spectrum-X Photonics system as a 512-lane, 200G-capable co-packaged switch in which multiple switch ASICs and their optical engines share one integrated assembly, including an on-package fiber "shuffle" that organizes the many fibers leaving the switch. [3][5]
The underlying silicon photonics comes from TSMC's COUPE (Compact Universal Photonic Engine) platform. COUPE uses 3D SoIC hybrid bonding to stack an electronic integrated circuit (EIC) directly on top of a photonic integrated circuit (PIC), minimizing the electrical path between the driver electronics and the optical devices. The photonic engines use micro-ring modulators (MRMs), compact ring-resonator structures that imprint data onto light far more area- and power-efficiently than the larger Mach-Zehnder modulators common in pluggable optics. [4][6][7]
A defining choice in NVIDIA's CPO design is that the lasers are not inside each optical engine. Instead, continuous-wave light is supplied from a smaller number of laser sources and routed to the modulators, which is what allows the "4 times fewer lasers" claim: fewer, more reliable laser sources replace the large population of individual lasers found across many pluggable modules. NVIDIA has noted that this continuous-wave laser source consumes on the order of 2 watts per port, with each optical engine drawing roughly 7 watts, and that removing pluggable optics yields about a 3.3 times reduction in the power attributable to the links themselves while improving signal quality by roughly 5.5 times (around 4 dB of loss in the co-packaged path versus about 22 dB through a conventional pluggable link). [4]
To make co-packaged optics serviceable in production, NVIDIA emphasized manufacturability: the optical engine is solder-reflow compatible and screened before attachment so that assembled units reach high yield, and the fiber connectors are detachable for surface-normal input and output, so optics can be connected and serviced without the whole package being a single non-repairable unit. The switches are liquid-cooled, both to handle the density of the silicon and to keep the on-package photonics at a stable temperature, which matters because micro-ring modulators are sensitive to temperature. [3][5]
NVIDIA's Spectrum-X Photonics family is delivered as liquid-cooled chassis based on the Spectrum-6 ASIC. The smaller SN6810 uses a single co-packaged Spectrum-6 device, while the larger SN6800 combines four ASICs in one system with an integrated fiber shuffle. [3][5]
| Switch | ASIC | Aggregate bandwidth | Port configuration | Cooling |
|---|---|---|---|---|
| Spectrum SN6810 | Spectrum-6 (single CPO device) | 102.4 Tb/s | 128 ports of 800 Gb/s (or 512 ports of 200 Gb/s) | Liquid-cooled |
| Spectrum SN6800 | Spectrum-6 (four CPO devices) | 409.6 Tb/s | 512 ports of 800 Gb/s (or 2,048 ports of 200 Gb/s) | Liquid-cooled |
NVIDIA's signaling uses 200G-class SerDes lanes (approximately 224 Gb/s per lane of raw signaling, with four lanes combining to form an 800 Gb/s port). [3][4]
In subsequent technical material focused specifically on the Ethernet platform, NVIDIA quantified the Spectrum-X Photonics benefits as up to a 5 times reduction in power per 1.6 Tb/s port compared with pluggable interconnects, 5 times longer flap-free (link-stable) AI uptime compared with off-the-shelf Ethernet, and 10 times greater network resiliency, by integrating the optical engines directly onto the switch ASIC. NVIDIA has also cited about 1.6 times greater bandwidth density for the Spectrum-X Photonics approach relative to traditional Ethernet, and roughly 1.3 times faster time to bring a deployment online. [3][5]
Spectrum-X Photonics and Quantum-X Photonics are two applications of the same co-packaged silicon-photonics technology aimed at the two dominant AI fabric types. Spectrum-X Photonics targets Ethernet and is built on the Spectrum-6 ASIC, while Quantum-X Photonics targets InfiniBand and is built on the Quantum-X switch silicon. The Quantum-X Photonics Q3450-LD packs four Quantum-X CPO sockets into a liquid-cooled chassis to deliver 144 ports of 800 Gb/s InfiniBand and about 115 Tb/s of aggregate bandwidth, using 200G SerDes and direct MPO fiber connectors, and it includes roughly 14.4 TFLOPS of in-network (SHARP-style) computing. [3][8]
| Platform | Fabric | Switch / model | ASIC | Aggregate bandwidth | 800 Gb/s ports | Availability |
|---|---|---|---|---|---|---|
| Spectrum-X Photonics | Ethernet | SN6810 | Spectrum-6 | 102.4 Tb/s | 128 | Second half of 2026 |
| Spectrum-X Photonics | Ethernet | SN6800 | Spectrum-6 | 409.6 Tb/s | 512 | Second half of 2026 |
| Quantum-X Photonics | InfiniBand | Q3450-LD | Quantum-X | 115 Tb/s | 144 | Second half of 2025 |
Spectrum-X Photonics complements, rather than replaces, the broader Spectrum-X Ethernet platform: it is the co-packaged-optics physical layer for that fabric, intended to be paired with NVIDIA's Ethernet SuperNICs and adapters for end-to-end AI networking. For customers who prefer InfiniBand, Quantum-X Photonics plays the equivalent role within NVIDIA's Quantum-X800 generation, which uses NVIDIA ConnectX adapters at the endpoints. [2][3]
NVIDIA stated that Spectrum-X Photonics and Quantum-X Photonics were developed with a broad silicon-photonics ecosystem spanning chip manufacturing, optical components, laser sources, connectors, and assembly. The partners named at the GTC 2025 announcement were TSMC, Browave, Coherent, Corning Incorporated, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric Industries, TFC Communication, Eoptolink, and Innolight. TSMC supplies the COUPE silicon-photonics process and SoIC 3D stacking; the others contribute optical engines, lasers, fiber and connector technology, and packaging and manufacturing capacity. [1]
At GTC 2025 NVIDIA said the Quantum-X Photonics InfiniBand switches would be available first, in the second half of 2025, with the Spectrum-X Photonics Ethernet switches following in 2026; NVIDIA's product materials specify the second half of 2026 for Spectrum-X Photonics, delivered through leading infrastructure and system vendors. The staggered schedule reflects NVIDIA shipping the technology into its InfiniBand line before its Ethernet line. [1][3]
Spectrum-X Photonics represents one of the first commercial moves of co-packaged optics from research demonstrations into mainstream AI networking products from a dominant vendor. Its importance lies less in raw switch bandwidth, which is comparable to NVIDIA's electrically connected Spectrum and Quantum switches, than in the power and reliability budget of the optics. As AI clusters scale toward hundreds of thousands and eventually millions of accelerators, the cumulative power drawn by pluggable optical modules and the failure rate of the lasers inside them become first-order constraints on how large and how dependable a fabric can be. By cutting interconnect power, reducing the number of lasers, and removing the per-link DSP retimers, NVIDIA argues that co-packaged optics frees power to support more GPUs within the same envelope and reduces the link disruptions that can stall large distributed training and inference jobs. The platform also marks a deepening of the NVIDIA and TSMC relationship into silicon photonics and 3D heterogeneous integration, and it has helped catalyze broader industry interest in co-packaged optics as a standard building block for future AI data centers. [1][4][5]