NVLink has gone through five generations as of the B200. Two separate version counters run in parallel and cause persistent confusion in vendor documentation: the NVLink protocol generation (what generation of NVLink each GPU implements) and the NVSwitch silicon generation (what chip sits inside the switch or the DGX node's onboard NVSwitch modules). These are different counters.
NVLink protocol generations:
NVLink 2 (Volta / V100, 2017): 300 GB/s total bidirectional per GPU, 6 links x 25 GB/s each. First generation that supported NVSwitch as an external switching element for all-to-all.
NVLink 3 (Ampere / A100, 2020): 600 GB/s total bidirectional per GPU, 12 links x 25 GB/s each. Doubled link count vs NVLink 2. DGX A100 shipped with 2nd-generation NVSwitch silicon at 7.2 Tb/s aggregate switch bandwidth.
NVLink 4 (Hopper / H100, 2022): 900 GB/s total bidirectional per GPU, 18 links x 25 GB/s each. 50% increase over NVLink 3. DGX H100 ships with 3rd-generation NVSwitch silicon at 13.6 Tb/s aggregate switch bandwidth -- covered in detail in Chapter 0 and revisited in Chapter 12.
NVLink 5 (Blackwell / B200, 2024): 1.8 TB/s total bidirectional per GPU -- double NVLink 4. 4th-generation NVSwitch silicon. The full DGX B200 configuration supports up to 72 GPUs in a single NVLink domain (up from 8 per DGX node in H100 architecture), connected by NVLink Switch modules.
The NVLink-C2C variant (used in GH200, covered in Act 6): NVLink Chip-to-Chip is a die-to-die interconnect variant running at 900 GB/s between the Grace CPU and the H100 GPU on the same package. It is not the same as the GPU-to-GPU NVLink -- it is a different electrical implementation optimised for short die-to-die distances with coherent memory access semantics.
The pattern across generations: each NVLink generation roughly doubles bandwidth, adding more links per GPU. The NVSwitch generation counter tracks the switch silicon and always lags one step behind the GPU -- the NVSwitch that ships inside a DGX node for a given GPU generation is the switch designed to aggregate that generation's link count.