Nvidia has launched its Vera Rubin chip which is designed to power orbital data centers and also enable workloads in space. The chip is part of a larger Pod Seven system, which consists of seven Vera Rubin GPUs across five rack-scale systems to create one AI supercomputer. The company also mentioned that this architecture is created to handle the unique demands of space-based computers, from satellite imagery analysis to autonomous spacecraft operations.
Despite the futuristic setting, many of the challenges remain the same as on Earth. Orbital data centers face issues of power efficiency, cooling, and reliability, all of which mirror terrestrial data center struggles. Nvidia CEO Jensen Huang acknowledged these hurdles but emphasized the company’s engineering talent: “We have engineers who can solve problems whether they’re on Earth or in orbit,” he said, underscoring Nvidia’s confidence in adapting its technology to extreme environments.
Nvidia’s broader strategy to maintain dominance in AI chips
The all new Vera Rubin chip launch clearly reflects Nvidia’s broader strategy to expand its dominance in AI chips into new domains. The spaced-based AI is seen as critical for realtime satellite data processing, defense applications and global communications. With the first-mover advantage Nvidia aims to capture demand from governments, aerospace firms, and also private space ventures.
This move from Nvidia also comes at a time when rivals like AMD and Intel are also exploring specialized chips for edge and aerospace computing.
Key features of the Vera Rubin chip and POD
Scale & Performance* 1,152 Vera Rubin GPUs across 40 racks
* 60 exaflops of compute power
* 10 PB/s bandwidth for ultra-fast data transfer
* Built on third-generation Nvidia MGX rack architecture
Specialized Rack Systems* NVL72 Compute Rack:
* 72 Rubin GPUs + 36 Vera CPUs
* Optimized for mixture-of-experts (MoE) and large context inference
* Up to 10x better inference performance per watt compared to Blackwell
* Groq 3 LPX Rack:
* 256 LPUs per rack for low-latency inference
* Designed for trillion-parameter models with long context lengths
* Delivers 35x more tokens and 10x more revenue opportunity vs. Blackwell
* Vera CPU Rack:
* 256 CPUs per rack
* Supports 22,500+ concurrent reinforcement learning environments
* Twice as efficient and 50% faster than traditional CPU racks
* BlueField-4 STX Storage Rack:
* AI-native storage with CMX context memory platform
* Offloads KV cache for massive context handling
* Provides 5x higher tokens-per-second and 5x better power efficiency
* Spectrum-6 SPX Networking Rack:
* Silicon photonics-based networking
* 102.4 Tb/s switch bandwidth with 200 Gb/s optics
* Ensures low latency, resiliency, and synchronized workloads
Energy & Cooling Innovations* Dynamic Power Steering: Allocates power to components that need it most
* Rack-Level Energy Storage: Capacitors smooth power swings, reducing peak current by 25%
* Intelligent Power Smoothing: Stabilizes workloads without massive battery packs
* Liquid Cooling at 45°C: Enables cost-efficient free cooling, boosting performance per watt
Scalability* NVL72 racks can scale up to NVL576 (576 GPUs) with all-to-all NVLink topology
* Future Kyber NVL1152 racks will double GPU domains, supporting extreme scale-up AI
Start a Conversation
Post comment