Nvidia's Blackwell Ultra GPU Sets New Efficiency Records

Why it matters

- The Blackwell architecture is built on a custom TSMC 4NP process, featuring 208 billion transistors and a dual-die chip connected by a 10 terabytes per second (TB/s) interconnect. The Blackwell Ultra version enhances this with 1.5 times more AI compute FLOPS and double the acceleration for attention layers compared to the standard Blackwell GPU. - In MLPerf benchmarks for inference, the GB300 NVL72 rack-scale system demonstrated a 45% performance increase over the preceding GB200 platform when running the DeepSeek R1 model. It showed up to five times the performance of the older Hopper GPU architecture in some tests. - Compared to the Hopper generation, the GB300 NVL72 delivers up

Key numbers

• Nvidia's new Blackwell Ultra GB300 GPU is establishing new performance benchmarks in long-context LLM inference, significantly outperforming its GB200 predecessor in speed and efficiency.
• The hardware is seen as enabling an "efficiency era" of AI, making contexts of over 100,000 tokens more viable and cost-effective for production systems.
• - The Blackwell architecture is built on a custom TSMC 4NP process, featuring 208 billion transistors and a dual-die chip connected by a 10 terabytes per second (TB/s) interconnect.
• The Blackwell Ultra version enhances this with 1.5 times more AI compute FLOPS and double the acceleration for attention layers compared to the standard Blackwell GPU.

Sources

• strategic importance
• The Blackwell architecture
• In MLPerf benchmarks

Quick answers