NVIDIA NVLink 6.0

PER LINK, UNIDIRECTIONAL. DERIVED, NOT PUBLISHED - NVIDIA prints no per-link rate for this generation. NVIDIA's official NVLink product page spec table (link-nvlink-switch-spec, accessed and re-verified 2026-10-05) confirms 3,000 GB/s per GPU and 36 max links per GPU for Sixth Generation NVLink: 3000 / 36 = 83.3 GB/s per link bidirectional, so 41.7 GB/s per link per direction - the recorded value. DIRECTION CONFIRMED IN NVIDIA'S OWN PROSE, which is the stronger evidence and was added to this record's basis on 2026-10-05: the same page defines the convention explicitly - 'NVLink is a 3 TB/s BIDIRECTIONAL, direct GPU-to-GPU interconnect' - and gives the prose form 'The sixth-generation NVLink enables 3 TB/s of bandwidth per GPU for the NVIDIA Vera Rubin platform - 1.7x more bandwidth than the previous generation and 12x the bandwidth of PCIe Gen6', plus 'The NVLink Switch interconnects every GPU pair at an incredible 3 TB/s' and 'NVIDIA Vera Rubin NVL72 ... for a total of 216 TB/s'. So 3,000 GB/s is a BIDIRECTIONAL PER-GPU AGGREGATE and the identity 36 x 41.67 x 2 = 3,000 closes exactly; 72 x 3 TB/s = 216 TB/s matches the published rack total. NVIDIA marks the table preliminary and subject to change. A secondary source (Wikipedia) states 3.6 TB/s bidirectional per GPU, but this figure does not appear in any first-party NVIDIA document and is not supported by NVIDIA's own spec page, so it is rejected rather than stored - see gotcha w7n-nvlink-6-3600-gbps-has-no-first-party-source.

Direct GPU-to-GPU links plus switch fabric; Vera Rubin NVL72 connects 72 GPUs all-to-all as one non-blocking compute fabric.

NVLink 6 Switch spans 8-GPU and 72-GPU domains. NVIDIA states the NVL72 rack interconnects every GPU pair at 3 TB/s and adds SHARP in-network reduction, plus control-plane resilience and hot-swappable switch trays.

Vera Rubin generation. HEADLINE CAPABILITY, the third required quantity for this generation: NVLink 6 doubles the LINK COUNT again, 18 -> 36, while the per-link rate per direction goes DOWN from 50 to 41.7 - so this is the first generation in the family where the per-GPU aggregate rose while the per-link rate fell, and it is the clearest demonstration that NVIDIA's scaling lever on this fabric has been link count far more often than SerDes rate. The stated capability is rack-scale: 72 GPUs as one non-blocking all-to-all compute fabric at 3 TB/s per GPU pair, 216 TB/s total, with 'control plane resilience, the ability to run with a partially populated rack, and hot-swapping of switch trays' as the generation's actual new feature set - those three operational properties, not the bandwidth, are what NVIDIA leads with on this generation. NVIDIA's own relative claim is '1.7x more bandwidth than the previous generation', consistent with 3000/1800 = 1.67, and '12x the bandwidth of PCIe Gen6'. NVIDIA's official NVLink product page spec table (accessed and re-verified 2026-10-05) confirms 3,000 GB/s per GPU and 36 max links per GPU for Sixth Generation NVLink, marked preliminary. The 3,600 GB/s figure from Wikipedia (link-wikipedia-nvlink) has no first-party NVIDIA support and is not used - the rejection is recorded as gotcha w7n-nvlink-6-3600-gbps-has-no-first-party-source rather than left implicit. Internal consistency check: 72 GPUs x 3 TB/s = 216 TB/s matches the NVL72 rack total, and 36 x 41.67 x 2 = 3,000 closes the per-GPU identity. ROUNDING NOTE, since the stored value is rounded to one decimal: 41.7 x 36 x 2 = 3,002.4 rather than exactly 3,000, because the exact per-link per-direction rate is 3000 / 36 / 2 = 41.667 GB/s. The 2.4 GB/s gap is 0.08% and is an artefact of the stored precision, not a different figure; a reader reproducing the identity should divide the vendor aggregate by the link count rather than multiply the stored rounded value back up. This is the only record in the family where the stored per-link rate needs more than one decimal to reproduce the published aggregate exactly. Inference consequence: tensor-parallel sharding wants the highest per-direction fabric rate available, because every layer boundary ships activations across the link; at 1,500 GB/s per direction per GPU against Vera Rubin's memory bandwidth the fabric is no longer the limiting term for a 72-GPU domain. PROVENANCE 2026-10-05: the per-link figure is DERIVED, and this is the generation where the direction label matters most - the prose calls the same 3 TB/s 'bidirectional' while the spec table prints a bare '3,000 GB/s' with no qualifier, so the table read alone invites exactly the 2x error this repo has been corrupted by twice. See also w7n-nvlink-per-gpu-aggregate-vs-per-link-rate.