NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance
NVIDIA’s Vera Rubin platform swept every MLPerf Training v6.1 benchmark and, in its first MLPerf Endpoints submission, achieved more than 10× generational throughput‑per‑watt gains for agentic inference. The NVL72 system delivered up to 2× higher performance per GPU than the prior GB300 NVL72 and scaled DeepSeek‑V3 from 256 to 1,024 GPUs with ~97 % efficiency, cutting training time from 16 min to ~4 min. These results showcase how NVIDIA’s CUDA‑powered stack, NVLink intra‑node networking, and Quantum‑X800 InfiniBand/Spectrum‑X Ethernet enable massive, power‑efficient scaling for large‑model training and agentic workloads. Engineers can now leverage Vera Rubin to accelerate model development and serve more AI agents per watt in production.
⚡ Key Takeaways
- Swept every MLPerf Training v6.1 benchmark and delivered >10× throughput‑per‑watt gains on MLPerf Endpoints for agentic AI.
- Uses NVIDIA CUDA‑powered software stack combined with NVLink intra‑node and Quantum‑X800 InfiniBand/Spectrum‑X Ethernet scale‑out networking.
- Achieved ~97 % scaling efficiency when expanding DeepSeek‑V3 from 256 to 1,024 GPUs, reducing training time from 16 min to ~4 min.
- NVL72 provides up to 2× more performance per GPU than the previous GB300 NVL72 system.
- Requires NVIDIA Vera Rubin hardware (NVL72, GB300/GB200) and the associated networking fabric; results are benchmark‑specific.
The dramatic per‑watt and scaling improvements let production teams train larger models faster and serve many more autonomous agents without proportionally increasing power costs, directly accelerating time‑to‑value for AI‑driven products.
✅ Practical Steps
- Deploy your training workloads on NVIDIA Vera Rubin NVL72 nodes using the CUDA‑powered software stack.
- Configure NVLink within each node and connect nodes via Quantum‑X800 InfiniBand or Spectrum‑X Ethernet to achieve near‑linear scaling.
- Validate agentic inference performance with MLPerf Endpoints workloads to measure throughput‑per‑watt gains.
Want the full story? Read the original article.
Read on NVIDIA Blog ↗