Nvidia unveils its most powerful GPU yet: 3nm Rubin for AI - hwcooling.net

At SIGGRAPH 2026, NVIDIA unveiled its latest AI accelerator, the Rubin GPU, built on advanced 3nm process technology and designed for data center and AI computing. The Rubin GPU consists of two interc... Read original →

NVIDIA's unveiling of the Rubin GPU marks a pivotal moment in AI chip evolution. Built on 3nm process technology with EUV, the chip features 336 billion transistors across two nearly 800mm² dies, delivering up to 25 PFLOPS in low-precision operations and 50 PFLOPS for inference. The dual-die architecture, supported by 896 Tensor Cores and 22TB/s memory bandwidth, signals a major leap in data center AI computing. This advancement pressures TSMC to scale production and improves downstream server vendors' integration capabilities, especially with PCIe 6.0 and NVLink adoption. Geopolitical tensions may force NVIDIA to restructure its supply chain amid U.S. export controls, particularly affecting critical components. Competitors like AMD and Intel are likely to accelerate their MI300 and Ponte Vecchio platforms to counter market dominance. Over the next 12–24 months, Rubin GPU will drive down data center compute costs, accelerate large model training, and stimulate upgrades in HBM4 and Transformer Engine technologies, reinforcing a self-reinforcing tech ecosystem.