A proprietary GPU-to-GPU interconnect with far higher bandwidth and lower latency than any network — the scale-up fabric inside a node.
What NVLink Is For
NVLink is a memory-semantic interconnect between accelerators. A GPU addresses a peer’s memory rather than sending it a message, and it does so at bandwidths no general-purpose network approaches. That is what allows a model to be sharded across the GPUs of a node as though they shared one address space.
It is a different layer of the problem from InfiniBand or Ethernet, and conflating the two is the most common mistake in cluster conversations. NVLink is the scale-up fabric; the cluster network is scale-out, and they do not compete.
The Numbers, and the Cliff
Per-GPU aggregate bandwidth has gone 600 GB/s in the third generation, 900 GB/s in the fourth, 1.8 TB/s in the fifth and 3.6 TB/s in the sixth, counting both directions as the vendor quotes them. A rack-scale domain of 72 accelerators reaches roughly 130 TB/s of in-domain bandwidth.
An 800G scale-out port is 100 GB/s each way, or 200 GB/s counted the same way as those NVLink figures. The step from inside the domain to outside it is therefore about an order of magnitude, and it is a cliff rather than a slope. Every decision about how a model is partitioned comes back to which side of that cliff a given exchange lands on.
Why It Is Not Optical
Scale-up links are extremely short and extremely wide — in-package, across a board, or through a passive backplane and cable cartridge within one rack. At those distances copper is cheaper, cooler and lower latency than any optical alternative, and the link count is high enough that per-lane cost dominates everything else.
That will not hold forever. As scale-up domains grow past a rack, the reach that passive copper can cover at 200 Gb/s per lane runs out — roughly a metre or two — and optical scale-up becomes an engineering question rather than a research one. That is precisely the pressure behind co-packaged optics and linear pluggable optics.
Where It Meets Your Cable Plant
You do not cable NVLink with transceivers, but you cable everything at its boundary. The edge of the scale-up domain is where the fabric adapters are, and each of them is a port: DAC or AOC for in-rack and in-row hops, transceivers over structured fibre for anything longer.
The practical consequence of larger NVLink domains is fewer, fatter, longer external links rather than more of them — and a shift in the mix from copper toward optics, because a rack-scale domain has already consumed all the distance copper was good for. DAC vs AOC vs transceiver cost puts numbers on that choice.
NVLink reference
| 3rd generation (A100) | 600 GB/s per GPU |
|---|---|
| 4th generation (H100) | 900 GB/s per GPU |
| 5th generation (Blackwell) | 1.8 TB/s per GPU |
| 6th generation (Rubin) | 3.6 TB/s per GPU |
| Rack-scale domain example | 72 GPUs, ~130 TB/s aggregate in-domain |
| Comparison, 800G fabric port | 100 GB/s each way, 200 GB/s counted both ways |
| Media | in-package and in-rack electrical, not pluggable optics |
| Semantics | memory-addressable, not packet-forwarded |
Related Terms
scale-up and scale-out, InfiniBand, GPUDirect RDMA, CPO, 1.6T OSFP-XD lane guide.