- Co-packaged optics
- An architecture that moves optical conversion onto the switch chip's own package instead of a pluggable module at the faceplate. By shortening the internal electrical path it cuts the power spent driving high-speed signals, and by integrating the many hot, failure-prone modules it shrinks the failure surface — at the cost of easy serviceability.
- Collective operation (all-reduce)
- A synchronized exchange in which every participating accelerator contributes a value, the values are combined (for all-reduce, summed), and the result is returned to every participant. No accelerator proceeds until all have finished, which makes training synchronous and hands the fabric control over cluster speed.
- Lossless fabric and RDMA
- Two requirements of the AI scale-out tier. A lossless fabric uses flow control to prevent packet drops rather than recover from them, keeping the tail short for synchronous collectives. RDMA (remote direct memory access) moves data directly between node memories without host-processor involvement, removing latency and CPU overhead.
- Memory semantics
- The property of a link (such as NVLink) that lets one accelerator directly read and write another's memory, addressing it almost as if local, without host involvement or explicit message-passing. It pools the group's memories into one larger effective resource — often the decisive factor in whether a model fits.
- Reticle limit
- The maximum area a lithography scanner can pattern onto a single die, which caps how large one processor can be. Because you cannot simply fabricate an enormous processor, scale-up interconnect wires many real processors together tightly enough to substitute for the larger one that manufacturing forbids.
- Scale-out fabric
- The longer-reach network that joins many scale-up domains into a cluster of thousands of accelerators. It offers reach at lower per-link bandwidth and higher latency than the scale-up tier, and is the domain of InfiniBand and Ethernet, where cost per bit and tail latency are set for most clusters.
- Scale-up domain
- A small, dense group of accelerators bound by very high-bandwidth, short-reach links into a single coherent machine, so their combined memory and compute act as one pooled resource. Realized by NVLink and NVSwitch; widening this domain keeps more expensive communication on the fast tier and raises utilization.
- Straggler
- Any lagging element — a slow link, congested path, marginal node, or failing transceiver — that stalls an entire synchronous step because all participants wait on the slowest. Straggler risk rises with cluster size, so the tail of the communication distribution, not the mean of compute, bounds usable throughput.