News
Nvidia Blackwell Ultra AI racks outperform GB200 NVL72 by 1.5x in DeepSeek model testing
Monday, February 16, 2026 at 05:20 PM
Nvidia Blackwell Ultra AI server racks demonstrate a 1.5x performance improvement in latency-sensitive tasks over the GB200 NVL72 when running DeepSeek open-source models.
Context
Nvidia’s recently deployed Blackwell Ultra AI racks, specifically the GB300 NVL72, have demonstrated a 1.5x performance advantage over the standard GB200 NVL72 in testing with DeepSeek’s latest reasoning models. This throughput lead is primarily driven by a 50% increase in HBM3e memory capacity to 288GB and enhanced FP4 tensor compute capabilities. The benchmarks highlight the platform's optimization for "test-time scaling," which allows advanced reasoning models to deliver higher accuracy during latency-sensitive inference workloads.
This hardware evolution marks a strategic shift toward "AI factories" optimized for agentic and physical AI. Following its release in H2 2025, the Blackwell Ultra architecture has solidified Nvidia’s lead in high-end inference by integrating the Dynamo software framework and advanced liquid-cooled designs. These enhancements allow data centers to process trillion-parameter models with significantly lower total cost of ownership, providing a critical performance bridge before the transition to the next-generation Vera Rubin architecture scheduled for late 2026.
Related Companies
Nvidia
NVDA