NVIDIA Unveils Next-Gen Vera Rubin Superchip with Two GPUs for 2024
NVIDIA is once again at the forefront of technological innovation with the unveiling of its next-generation Vera Rubin Superchip. As industries increasingly turn to artificial intelligence (AI) for transformative solutions, this new superchip promises to be a game-changer, paving the way for more sophisticated AI applications and data processing capabilities.
NVIDIA's Vera Rubin Superchip: A Glimpse into the Future of AI
During the recent GTC event in Washington, NVIDIA's CEO Jensen Huang presented the Vera Rubin Superchip for the first time. This advanced architecture is designed with formidable power, featuring a combination of the Vera CPU and two Rubin GPUs.
What sets this superchip apart is its impressive memory configuration. The motherboard supports 32 sites of LPDDR system memory, which will work in tandem with the high-bandwidth memory (HBM4) integrated into the Rubin GPUs. This strategic design allows for enhanced data throughput, essential for demanding AI tasks.
The Rubin GPUs are currently in the lab, reflecting NVIDIA's aggressive timeline for development. They are among the first chips produced at TSMC, a leader in semiconductor fabrication. Each GPU employs multiple power circuitry setups and will boast 8 HBM4 memory sites along with dual reticle-sized GPU dies. Additionally, the Vera CPU will incorporate 88 custom ARM cores, allowing for a remarkable 176 threads of execution.
As for the production timeline, Jensen announced ambitious plans to begin mass production of the Rubin GPUs by late 2026, potentially by Q3 or Q4. This aligns with the ongoing rollout of NVIDIA's Blackwell Ultra "GB300" Superchip platforms, demonstrating the company's commitment to staying ahead in AI technology.
Specifications and Performance of the Vera Rubin NVL144 System
The Vera Rubin NVL144 platform is set to make waves in the AI sector with its cutting-edge specifications. It will utilize two reticle-sized Rubin GPUs, achieving an astounding 50 PFLOPs of FP4 performance, complemented by a total of 288 GB of next-generation HBM4 memory.
- Vera CPU: 88-core architecture with custom ARM design
- Threads: 176 simultaneous threads
- Interconnect: Up to 1.8 TB/s via NVLINK-C2C
This platform is not just about raw performance; it boasts remarkable scaling capabilities as well. The Vera Rubin NVL144 will deliver 3.6 Exaflops of FP4 inference and 1.2 Exaflops for FP8 training, representing a significant leap over previous generations, particularly the GB300 NVL72.
Comparative Performance Metrics
When comparing the Vera Rubin NVL144 to its predecessors, the advancements are striking:
- Performance Increase: 3.3 times more than GB300 NVL72
- Memory Bandwidth: 13 TB/s of HBM4 memory
- Fast Memory: 75 TB capacity
- NVLINK and CX9: Doubled capabilities at 260 TB/s and 28.8 TB/s, respectively
Introducing the NVIDIA Rubin Ultra NVL576 System
Further advancing NVIDIA's offerings, the Rubin Ultra NVL576 system is projected for release in the second half of 2027. This platform will expand on the capabilities of the NVL144, scaling from 144 to an impressive 576.
Key features of the Rubin Ultra include:
- GPU Configuration: Four reticle-sized chips
- Performance: Up to 100 PFLOPs of FP4
- Memory Capacity: 1 TB of HBM4e distributed across 16 HBM sites
Performance scaling is equally impressive, with the Rubin Ultra NVL576 expected to deliver:
- FP4 Inference: 15 Exaflops
- FP8 Training: 5 Exaflops
- Overall Memory Bandwidth: 4.6 PB/s of HBM4 memory
- Fast Memory: 365 TB capacity
This architecture will offer a staggering increase in performance capabilities over its predecessor, exhibiting a 14x improvement compared to the GB300 NVL72.
NVIDIA's Comprehensive AI GPU Roadmap
NVIDIA's roadmap for its data center and AI GPUs showcases a clear trajectory of innovation and development. Below is a summarized comparison of the upcoming GPU families:
| GPU Codename | Feynman | Rubin (Ultra) | Rubin | Blackwell (Ultra) | Blackwell | Hopper | Ampere | Volta | Pascal |
|---|---|---|---|---|---|---|---|---|---|
| GPU Family | GF200? | GR300? | GR200? | GB300 | GB200/GB100 | GH200/GH100 | GA100 | GV100 | GP100 |
| GPU SKU | F200? | R300? | R200? | B300 | B100/B200 | H100/H200 | A100 | V100 | P100 |
| Memory | HBM4e/HBM5? | HBM4 | HBM4 | HBM3e | HBM3e | HBM2e/HBM3/HBM3e | HBM2e | HBM2 | HBM2 |
| Launch | 2028 | 2027 | 2026 | 2025 | 2024 | 2022-2024 | 2020-2022 | 2018 | 2016 |
NVIDIA continues to evolve its product line, focusing on meeting the growing demands of AI and data processing. As these superchips progress from concept to reality, the implications for industries ranging from healthcare to finance could be profound, enabling unprecedented levels of analytics, automation, and predictive capabilities.
For those interested in a deeper dive into NVIDIA's upcoming technologies, you can check out this insightful video:
Leave a Reply