NVIDIA's Vera Rubin NVL72 System Achieves 30x Throughput Boost for AI Workloads!

At the AI Infra Summit, NVIDIA showcased its Vera Rubin NVL72 system achieving 30 times higher throughput per megawatt than its GB300 counterpart for agentic coding tasks. This benchmark, revealed alongside power constraints impacting AI revenues, utilized real session data. The new AgentX method, featuring 393 coding traces, emphasizes real-world efficiency. Innovations like disaggregated serving and large-scale expert parallelism drive its performance. The results, which are still subject to software optimization, may herald a shift in AI energy metrics.