Nvidia Roadmap — Single Machine
| SuperPod System | Generation | Accelerator | Form Factor | # of Logical GPUs | # of GPU dies | NVLink Generation | Scale-Up Topology | Within Rack Scale-Up | Between Rack Scale-Up | Bandwidth per Logical GPU (GB/s uni-d) | Lane Speed (GB/s) | NVLink Switch Generation | Switch Aggregate BW (GB/s uni-d) | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Rubin NVL8 HGX | 2026 | Rubin | HGX | 8 | 16 | NVLink 6 | One-Layer Switched All-to-All | UBB (PCB) | N/A | 1,800 | 400G (200G PAM4 Simultaneous Bidi) | NVLink 6 Switch | 3,600 |