NVIDIA ships the first batch of Rubin GPUs, replacing copper NVLink with photonic interconnects for 10x bandwidth gains. See the hardware engineering.
Why NVIDIA Moved From Copper to Light
For years, the links between GPUs relied on copper traces carrying electrical signals. Copper works well over short distances, but as bandwidth demands climb, it runs into physics: electrical signals attenuate quickly, generate heat, and require more power to push data across a board or between chassis. Rubin swaps that copper NVLink fabric for photonic interconnects, moving data as light rather than electrical current. That single change targets a roughly tenfold jump in bandwidth between GPUs.
The reason light helps is straightforward. Optical signals hold their integrity over longer distances with far less loss, and they sidestep much of the electromagnetic interference that limits how fast and how far copper can carry a signal. For a chip whose usefulness depends on how quickly it can talk to its neighbors, the interconnect is often the real bottleneck, not the compute cores themselves.
What Photonic Interconnects Actually Change
The headline number is bandwidth, but the more interesting consequences are structural. When the fabric between GPUs stops being the limiting factor, you can build larger coherent clusters that behave more like a single large accelerator. Workloads that were previously split awkwardly across nodes — because moving activations or gradients between them was too slow — can be scheduled with less concern for where each piece of data physically lives.
- Longer reach: optical links let GPUs sit farther apart without a bandwidth penalty, loosening the tight physical packing that copper forces.
- Lower signaling overhead: less energy spent driving signals across the interconnect leaves more of the power budget for actual computation.
- Larger effective memory pools: faster GPU-to-GPU transfer makes it more practical to treat many GPUs' memory as one shared space.
Engineering Tradeoffs to Watch
Photonics is not a free win. Converting between electrical and optical domains adds components — lasers, modulators, and photodetectors — that introduce their own failure modes and manufacturing complexity. Optical alignment is unforgiving, and the packaging that puts light sources near the compute die is harder to build and test than laying down copper. Early batches of any new interconnect technology tend to expose these integration challenges before yields and reliability settle.
There is also a systems question. A tenfold bandwidth increase between GPUs only pays off if the rest of the stack — memory, networking beyond the local fabric, and the software scheduling work across the cluster — can keep the pipes full. Faster interconnects shift the bottleneck rather than eliminate it, so the practical gains depend heavily on the specific workload.
What This Means for People Building on It
If you run large distributed training or inference, the useful mental model is that the cost of communication between GPUs is dropping sharply relative to the cost of computation. That tends to favor model and parallelism strategies that were previously communication-bound: more aggressive tensor and pipeline parallelism, and larger batches spread across more devices. It is worth revisiting assumptions baked into existing code about how expensive cross-GPU transfers are.
For anyone evaluating early Rubin hardware, treat the first batch as exactly that — early. Validate that your framework and drivers actually exercise the photonic path, benchmark your own workloads rather than trusting the interconnect's peak numbers, and watch for the integration rough edges that any first-generation optical fabric is likely to carry.