Photonics for GPUs
At the Montgomery Summit, panels highlighted that GPUs—originally built for graphics—have become the backbone of AI, and NVIDIA is pushing photonics with Spectrum‑X and Quantum‑X switches that it says cut power use ~3.5x and raise resiliency about 10x (x.com). The takeaway is practical: optical switching and co‑packaged optics are being framed not as replacements for GPUs but as complementary tech that slashes interconnect power as AI models scale (x.com).
The strangest part of the artificial intelligence boom is that the most expensive machines in the room are not always the ones setting the next limit. The graphics processing units do the math, but the wires between them are starting to decide how big an artificial intelligence system can get. (nvidianews.nvidia.com) A graphics processing unit began life as a chip for drawing video game scenes. It became the core engine for artificial intelligence because the same design that handles millions of pixels at once is also good at handling millions of matrix calculations at once. (developer.nvidia.com) That shift changed the shape of the data center. A modern artificial intelligence training cluster is no longer one powerful server but a giant team of servers that must act like one machine while they split a model across many graphics processing units. (developer.nvidia.com) Once a model is spread across hundreds or thousands of chips, the network stops being a side detail. Every delay in moving data from one graphics processing unit to another can leave expensive processors waiting idle, like delivery trucks lined up at a warehouse with the loading dock blocked. (nextplatform.com) That network is also becoming a power problem. As artificial intelligence clusters grow, the electricity used to push signals across racks and buildings rises fast enough that interconnect power has become a design constraint, not just an operating expense. (optics.org) (datacenterdynamics.com) This is where photonics enters the story. Instead of relying only on electrical signals moving through copper traces and pluggable modules, photonics uses light to carry data, which helps over longer distances and can reduce the energy wasted in the handoff between chip and network link. (photonics.com) (optics.org) The key phrase is co-packaged optics. That means the optical components are moved right next to the switch silicon instead of sitting farther away in separate pluggable transceivers, which cuts the number of electrical hops before data becomes light. (nvidianews.nvidia.com) (convergedigest.com) NVIDIA is now trying to make that architecture mainstream around its graphics processing unit clusters. On March 18, 2025, the company announced Spectrum-X Photonics for Ethernet and Quantum-X Photonics for InfiniBand, both aimed at linking artificial intelligence systems at very large scale. (nvidianews.nvidia.com) The company’s pitch is not that light replaces the graphics processing unit. The pitch is that optical switching and co-packaged optics remove a growing bottleneck around the graphics processing unit by making the network cheaper to power and easier to scale. (nvidianews.nvidia.com) (nextplatform.com) NVIDIA says these switches deliver 1.6 terabits per second per port and can connect millions of graphics processing units across sites. It also says the design uses 4 times fewer lasers and delivers 3.5 times better power efficiency, 63 times greater signal integrity, 10 times better network resiliency at scale, and 1.3 times faster deployment than traditional methods. (nvidianews.nvidia.com) (photonics.com) Those numbers are company claims, and outside coverage has noted that NVIDIA did not publish full benchmark details in the announcement. Even so, the direction is clear: the company is treating the network fabric around artificial intelligence chips as a product category important enough to redesign from the physics up. (datacenterdynamics.com) (nvidianews.nvidia.com) That is why the Montgomery Summit discussion landed on photonics as a practical story, not a science project. If artificial intelligence models keep expanding, the next gains may come less from replacing the chip that does the thinking and more from rebuilding the roads that let all those chips think together. (nvidianews.nvidia.com)
