NVIDIA Launches AI Space Computing

🔑 Enhanced Key Takeaways

• •NVIDIA Vera Rubin GPU features a 2.3 kW TDP and 50 petaFLOPS of NVFP4 AI inference performance per GPU, with 288GB HBM4 capacity.
• •Vera CPU includes 88 custom Olympus Arm cores supporting spatial multithreading for 176 threads, paired with up to 1.5TB LPDDR5X memory at 1.2 TB/s bandwidth.[1][2][3][4][5]
• •Rubin platform introduces sixth-generation NVLink at 3.6 TB/s per GPU and NVLink-C2C at 1.8 TB/s for coherent CPU-GPU memory access.[1][3][4]
• •Vera Rubin NVL72 rack delivers 3.6 exaFLOPS NVFP4 inference and integrates components like NVLink 6 switch, ConnectX-9, and BlueField-4 DPU.[2][3][4][6]

🛠️ Technical Deep Dive

• •Rubin GPU: 224 Streaming Multiprocessors (SMs) with fifth-generation Tensor Cores for NVFP4/FP8, expanded SFUs for attention/sparse compute; 288GB HBM4 (8 stacks), 22.2 TB/s bandwidth, 2.3 kW TDP.[1][3][4]
• •Vera CPU: 88 Olympus Armv9.2 cores, spatial multithreading (176 threads/core via resource partitioning), <50W for 1.5TB LPDDR5X at 1.2 TB/s bandwidth using SOCAMM modules.[2][3][4][5]
• •Interconnects: NVLink 6 at 3.6 TB/s GPU-to-GPU, NVLink-C2C 1.8 TB/s coherent CPU-GPU; NVL72 rack: 72 GPUs + 36 Vera CPUs, 20.7TB HBM4, 260 TB/s aggregate NVLink.[1][2][3][4]
• •Rack-scale: NVL72 offers 3.6 EFLOPS NVFP4 inference; includes BlueField-4 DPU (64-core Grace CPU), ConnectX-9 networking.[2][3][4][6]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline