AWS NeuronLink (Neuron chip-to-chip scale-up)

The scale-up fabric for both Inferentia2 and Trainium2 in this repo. AWS states 192 GB/s NeuronLink on the Inf2 page and does not publish a distinct NeuronLink rate for Trainium2, so the same figure is carried for both parts rather than inventing a Trainium2-specific number. CONTEXT THAT SETS THE SCALE: the 192 GB/s is roughly 40-45% of NVLink 4's 450 GB/s per direction on an H100 SXM, and roughly a fifth of the bidirectional aggregate a reader would compare it against if they ignored this record's basis. So on raw fabric bandwidth Neuron is well behind NVSwitch for tensor parallelism at high degree - and the reason AWS parts still work for large-model inference is capacity and the chip count available per instance, not fabric speed: 12 chips x 32 GB on Inf2, 16 chips x 96 GB on Trn2.48xlarge. THAT IS THE INFERENCE RELEVANCE, and it is a capacity-first rather than bandwidth-first argument: with a slow fabric you want the fewest chips that hold the model, and you want pipeline parallelism (which needs only point-to-point transfers per stage boundary, not all-to-all) over wide tensor parallelism (which needs all-reduce bandwidth at every layer). AWS's own framing supports this - the Inf2 page justifies NeuronLink specifically 'for ultra-large models that do not fit into a single chip'. Two independent Neuron notes for a reader who confuses them: NeuronLink is chip-to-chip INSIDE an instance, and EFA (see aws-efa) is the scale-out network BETWEEN instances and hosts; conflating them is the single most common error with AWS interconnect diagrams. Also note the compiler story interacts here - Neuron autocasts FP32/FP16 models down to BF16/FP8/cFP8 at compile time, which reduces the bytes that have to cross this link, and is a cheaper lever than the fabric itself. KIND: filed as 'other' because NeuronLink has no entry in the schema's interconnect enum and is not a member of any named family.