SANTA CLARA — At a packed developer keynote at NVIDIA headquarters, Chief Executive Jensen Huang unveiled the Blackwell Ultra B300 GPU accelerator, a hardware powerhouse purpose-built to eliminate memory capacity bottlenecks during extended test-time reasoning.
Featuring a staggering 288 gigabytes of ultra-high-bandwidth HBM3e memory operating at 8.0 terabytes per second of aggregate bandwidth, the B300 allows an entire 70-billion-parameter model to reside entirely in high-speed VRAM on a single dual-die package, completely bypassing NVLink interconnect latencies for inference operations.
Tackling the Reasoning Memory Wall
As modern artificial intelligence models transition from single-turn retrieval to continuous multi-step reasoning, key-value (KV) cache storage requirements have exploded. Running multiple concurrent agent reasoning threads quickly consumes tens of gigabytes of working memory per active user session.
“You cannot scale autonomous agent reasoning if your GPU is spending half its clock cycles waiting for memory pages to swap across the bus,” Huang proclaimed, holding up a liquid-cooled B300 dual-package compute tray. “Blackwell Ultra turns an entire rack into a coherent single-GPU compute entity.”
Hyperscale cloud providers, including Microsoft Azure, Amazon Web Services, and Google Cloud, have committed to deploying initial B300 clusters beginning early next quarter.