NVIDIA's Vera Rubin Chips Claim 30 Times More Tokens Per Watt Than GB300
Ian Buck told Santa Clara's 8,000-strong crowd, more than double last year's, that DSX MaxLPS adds 40% more GPU capacity per megawatt.
- Vera Rubin NVL72 alone hits 30X GB300's throughput per megawatt on the DeepSeek V4 Pro model.
- Paired with the new Groq 3 LPX chip, the platform claims 35X GB200's throughput per megawatt.
- SemiAnalysis's AgentX benchmark found a 45X lower cost per million tokens than current rigs.
Why it matters: A new chip outrunning last year's chip is not news; it is the entire business model.