NVIDIA Quantum-X Photonics
Last reviewed
Jun 3, 2026
Sources
12 citations
Review status
Source-backed
Revision
v2 ยท 2,266 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
12 citations
Review status
Source-backed
Revision
v2 ยท 2,266 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA Quantum-X Photonics is a family of co-packaged-optics network switches that NVIDIA announced at its GTC conference on March 18, 2025. The switches carry the InfiniBand protocol and they fold the optical engines that convert electrical signals into light directly into the switch package, next to the switching chip, instead of relying on the pluggable transceiver modules that data-center switches have used for years. NVIDIA pitched the design as a way to wire together the very large GPU clusters it calls AI factories, where the older approach to optics starts to cost too much power and fail too often. A sibling product, NVIDIA Spectrum-X Photonics, applies the same idea to Ethernet. [1][2]
The two families were the first co-packaged-optics switches NVIDIA brought to market, and they sit alongside the company's existing networking gear rather than replacing it. Quantum-X Photonics targets the InfiniBand fabrics that connect GPUs inside training clusters, while Spectrum-X Photonics extends the same optics into NVIDIA's Spectrum-X Ethernet line. Both were built with a group of silicon, optics, and packaging partners led by TSMC. NVIDIA later set commercial availability for the Quantum-X InfiniBand switches at early 2026, with the Spectrum-X Ethernet switches following in the second half of 2026. [1][3][11]
In a conventional switch, the switching chip, or ASIC, sits in the middle of a circuit board. To reach an optical fiber, an electrical signal has to leave the ASIC, run across the board, pass through a connector, and travel into a pluggable transceiver plugged into the front panel. The transceiver is the part that turns electrons into photons. Only there does the signal become light and head out over the fiber. Each hop along that electrical path loses signal and burns power, and a big switch has hundreds of these front-panel modules. [4]
Co-packaged optics, usually shortened to CPO, collapses that chain. The electrical-to-optical conversion moves onto the switch package itself, so the optical engine sits right beside the ASIC and the fiber plugs almost directly into it. The electrical signal no longer has to survive a long trip across the board and through a stack of connectors before it becomes light. NVIDIA puts the signal loss for a 200 gigabit-per-second channel at about 22 decibels with pluggable optics and roughly 4 decibels with its co-packaged design. Lower loss means the chips driving each link can be simpler and can run cooler. [4]
The reason this matters for AI is scale. A cluster with tens or hundreds of thousands of GPUs needs a matching number of high-speed optical links, and pluggable transceivers are both power-hungry and failure-prone in those quantities. NVIDIA cites roughly 30 watts per pluggable interface against as little as 9 watts for its co-packaged version. Multiply that gap across a network that spans a whole building and the savings reach into the megawatts, which is power that can go to GPUs instead. The sheer count of discrete modules and the connectors they plug into is also a reliability problem, because every module and every seated connector is one more thing that can drift or drop a link. By integrating the optics, NVIDIA says it cut the laser count by a factor of four and improved network resiliency at scale by about ten times. [1][4][5]
NVIDIA framed the headline gains for the photonics switches as 3.5 times better power efficiency, 63 times greater signal integrity, 10 times better network resiliency at scale, and 1.3 times faster deployment, all measured against the traditional pluggable approach. The 1.3 figure reflects the fact that there are fewer parts to install and fewer links to bring up. The company estimated that for a site on the order of 100,000 servers, moving from pluggable optics to co-packaged optics could drop the optics power draw from around 40 megawatts to about 21.6 megawatts, and it described the freed-up budget as room for roughly three times the GPUs in the same optics power envelope. [1][5]
Quantum-X Photonics is the InfiniBand member of the pair, aimed at the back-end fabric that links GPUs during training. Each switch provides 144 ports of 800 Gb/s InfiniBand built on 200 Gb/s SerDes, and it uses a liquid-cooled design to pull heat off the onboard silicon photonics. NVIDIA describes the lead platform as the Q3450-LD, also reported as the Quantum 3450-LD, which packs four Quantum-X co-packaged sockets into one chassis, wired together in a non-blocking arrangement to deliver those 144 ports at 800 Gb/s for about 115 Tb/s of aggregate bandwidth. The switch also carries NVIDIA's in-network computing through the fourth generation of its SHARP protocol, which offloads parts of collective operations like all-reduce onto the switch itself. [1][3][6]
Against the previous InfiniBand generation, NVIDIA claims Quantum-X Photonics offers about two times the speed and five times the scalability for AI compute fabrics. The liquid cooling is not optional flavor here. The optical engines now share a package with a hot switching ASIC, so keeping the lasers and modulators in a stable thermal window matters more than it did when the optics lived in separate front-panel modules. [1][6]
Spectrum-X Photonics brings co-packaged optics to Ethernet, extending NVIDIA's Spectrum-X networking platform. It comes in more than one configuration. NVIDIA describes a model in the SN6810 class with 128 ports of 800 Gb/s for about 102.4 Tb/s of aggregate bandwidth, and a larger SN6800-class system with 512 ports of 800 Gb/s, or 2,048 ports of 200 Gb/s, for roughly 409.6 Tb/s of total bandwidth across four ASICs. The press materials summarized the range as a 100 Tb/s option and a 400 Tb/s option. The bigger switch uses an integrated fiber shuffle inside the quad-ASIC package, which lets a single large cluster scale flat without an extra layer of cabling. [1][7][8]
NVIDIA positions Spectrum-X Photonics as the Ethernet on-ramp for operators who want co-packaged optics but run their AI fabric on Ethernet rather than InfiniBand. The company quotes a power reduction of about five times per 1.6 Tb/s port against pluggable interconnects, and roughly five times longer flap-free uptime compared with off-the-shelf Ethernet, where a link flap is a connection that drops and re-establishes and can disrupt a training run. NVIDIA later restated the same gains as 5 times better power efficiency, 5 times longer sustained AI application runtime, and 1.3 times faster time to insight against pluggable transceivers, and it described Spectrum-X Photonics as the flagship switch for the networking of its Vera Rubin generation of AI infrastructure. [7][8][11]
| Item | Quantum-X Photonics | Spectrum-X Photonics |
|---|---|---|
| Protocol | InfiniBand | Ethernet |
| Lead platform / models | Q3450-LD (Quantum 3450-LD) | SN6810, SN6800 |
| Port configuration | 144 ports at 800 Gb/s | 128 ports at 800 Gb/s (SN6810); 512 ports at 800 Gb/s or 2,048 ports at 200 Gb/s (SN6800) |
| Aggregate bandwidth | About 115 Tb/s | About 102.4 Tb/s (SN6810); about 409.6 Tb/s (SN6800) |
| SerDes | 200 Gb/s | 200 Gb/s |
| In-network compute | SHARP, fourth generation | NVIDIA Spectrum-X features |
| Cooling | Liquid-cooled | Liquid-cooled |
| Stated availability | Early 2026 | Second half of 2026 |
Sources: NVIDIA newsroom and product pages, with model-level detail reported by The Next Platform and Tom's Hardware. [1][3][6][7][11]
The optics inside these switches come from a silicon-photonics process rather than from discrete optical components glued onto a board. NVIDIA built the engines on TSMC's platform called COUPE, short for Compact Universal Photonic Engine. COUPE uses TSMC's SoIC 3D stacking to bond a 65 nanometer electronic integrated circuit, the chip that drives the optics, face to face onto a photonic integrated circuit, the chip that carries the light. Stacking the two shortens the electrical path between the driver and the optical modulator to almost nothing, which is a large part of where the power and signal-integrity gains come from. [9][10]
The design leans on micro-ring modulators to encode data onto the light. A micro-ring is a tiny looped waveguide that is more compact and lower power than the Mach-Zehnder modulators often used in earlier optics, which helps pack many channels into a small package. The lasers themselves stay outside the main package as external laser sources, so the hottest, most failure-sensitive parts are kept away from the switching ASIC and can be serviced on their own. NVIDIA also uses detachable fiber connectors so that fibers can be attached and replaced in the field rather than soldered down. [9][10]
Beyond TSMC, NVIDIA credited a wider group of partners for the silicon, the optical process, and the supply chain, including Browave, Coherent, Corning, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric Industries, and TFC Communication. The breadth of that list reflects how many separate disciplines a co-packaged-optics switch pulls together, from wafer fabrication to laser supply to fiber and connector work. [1]
NVIDIA describes the current era of AI data centers as AI factories, buildings whose product is intelligence and whose raw material is electricity. The pitch for Quantum-X Photonics and Spectrum-X Photonics is that networking power, not just compute power, becomes a real ceiling once a site reaches hundreds of thousands or a million accelerators. Every GPU added to a training job needs fast links to the others, and with pluggable optics the transceivers alone can soak up a sizable slice of the facility's power and become a steady source of link failures. Co-packaged optics is NVIDIA's answer to both problems at once. [1][5]
The photonics switches handle the longer reaches inside and between racks, the scale-out fabric, while NVLink handles the dense, short-reach connections among GPUs inside a rack, the scale-up fabric. Quantum-X Photonics carries the InfiniBand version of that scale-out fabric and Spectrum-X Photonics carries the Ethernet version, so an operator can pick the protocol that fits and still get the power and reliability benefits of integrated optics. Jensen Huang tied the launch to that ambition, saying that by integrating silicon photonics directly into switches NVIDIA was opening the gate to million-GPU AI factories. [1]
When it announced the switches at GTC 2025, NVIDIA had said the Quantum-X Photonics InfiniBand switches would arrive later in 2025 and the Spectrum-X Photonics Ethernet switches in 2026. The company later moved the InfiniBand date out, setting commercial availability for the Quantum-X InfiniBand switches at early 2026, with the Spectrum-X Ethernet switches following in the second half of 2026 from infrastructure and system vendors. The staggered timing keeps the InfiniBand variant first, which fits its role in the training clusters that drove the early demand, with the Ethernet variant following for operators standardizing on Ethernet. [1][3][7][11]
The first named deployments came at the SC25 supercomputing conference in November 2025, where NVIDIA said cloud and supercomputing operators including Lambda and CoreWeave would adopt Quantum-X Photonics co-packaged-optics switches, alongside a deployment at the Texas Advanced Computing Center. For the Ethernet side, NVIDIA's silicon-photonics materials list CoreWeave, Lambda, Meta, Microsoft, and Oracle Cloud Infrastructure among the early adopters lining up Spectrum-X Photonics for Vera Rubin AI factories. [11][12]
Co-packaged optics had been discussed in the networking industry for years before NVIDIA's announcement, often as a promising idea that kept running into manufacturing and serviceability hurdles. Bringing it into shipping InfiniBand and Ethernet switches aimed at the largest AI buildouts gave the approach a concrete, high-volume use case, and it signaled that the power cost of moving data, not only of doing math, had become a first-order constraint in AI infrastructure. By late 2025 industry coverage was describing AI networking as the technology's first real high-volume application, with NVIDIA and Broadcom both pushing co-packaged-optics switches toward production. [4][5][12]
The approach also comes with tradeoffs. Integrating optics into the switch package makes the switch harder to service than one with hot-swappable front-panel modules, which is part of why NVIDIA leaned on external laser sources and detachable fiber connectors to keep the failure-prone pieces replaceable. The switches need liquid cooling, which assumes a facility built or retrofitted for it. The efficiency and resiliency figures come from NVIDIA's own comparisons against pluggable optics rather than from independent testing, so they are best read as the vendor's claims at launch. And the broadest benefits land at very large scale, where the link and power counts are high enough for integrated optics to pay off, rather than in small deployments. [1][4]