Key Takeaways
1. Nvidia has taped out six unique Rubin architecture chips, currently being tested at TSMC for trial production.
2. The redesign includes CPUs, various GPU versions, an upgraded NVLink switch, networking silicon, and a silicon-photonics processor, marking a significant architectural overhaul.
3. The Rubin R100 GPUs will feature next-generation HBM4 stacks, improving bandwidth and power delivery compared to HBM3E.
4. The Rubin platform is designed to meet the growing demands of large data centers for AI workloads, with enhanced software tools for developers.
5. Nvidia plans to launch the Rubin chip family around 2026, with Rubin Ultra following in the next year, depending on TSMC’s production readiness.
Nvidia’s chief executive, Jensen Huang, shared exciting news during his recent trip to Taiwan. He announced that the company has successfully taped out six unique Rubin architecture chips. Currently, these chips are at TSMC, where they are being tested and getting ready for trial production.
Major Architecture Changes
This development signifies a complete redesign of the architecture, moving beyond the usual small GPU updates. Instead of merely enhancing the GPU, this overhaul includes CPUs, various GPU versions, an upgraded NVLink switch, networking silicon, and a silicon-photonics processor to allow for optical connections at the rack level. For the first time, Nvidia will use a chiplet design, taking advantage of TSMC’s cutting-edge 3nm N3P process node with CoWos-L packaging. The architecture is designed with a larger 4x reticle, which is an increase from the Blackwell’s 3.3x reticle size.
Advanced Features and Performance
The new Rubin R100 GPUs are set to feature next-generation HBM4 stacks, with specially designed base dies to support enhanced bandwidth and power delivery at larger scales compared to the current HBM3E. This new platform is anticipated to be a significant advancement, similar to the leap Hopper made in computational power. Testing for thermal performance, power usage, and interconnect efficiency has already begun.
Future Plans and Launch Timeline
The Rubin platform is aimed at satisfying the increasing demands of large data centers as AI workloads continue to grow. Nvidia is also enhancing its software tools, enabling developers to utilize the new features and connections immediately.
Nvidia is targeting a launch for the Rubin chip family around 2026, with Rubin Ultra following in the subsequent year. However, this timeline is contingent on TSMC’s production capabilities and readiness. During his visit, Huang expressed gratitude towards TSMC staff for their critical contribution to the Rubin project.
Source:
Link
Leave a Reply