Overview
- The engineering sample integrates a custom 88-core Arm-based Vera CPU alongside two Rubin GPU packages on a single board for AI and HPC workloads.
- Event imagery indicates each Rubin package uses two compute chiplets with eight HBM4 stacks, and the Vera CPU shows a multi‑chiplet design.
- Jensen Huang cited roughly 100 petaFLOPS of FP4 AI performance and claimed the platform contains no fewer than six trillion transistors.
- Markings on the demo hardware show the Rubin GPU packages were assembled in Taiwan in the 38th week of 2025, suggesting testing units are already in labs.
- Nvidia says Rubin GPUs and the Vera CPU are slated for production around Q3–Q4 2026, with the NVL144 platform in H2 2026 and a larger Rubin Ultra NVL576 system following in 2027.