News

AMD introduces Helion rack architecture using 2nm and 3nm nodes to compete with NVIDIA Rubin

Monday, December 8, 2025 at 06:31 PM

AMD has introduced its Helion rack architecture designed to compete with NVIDIA's Rubin. The system utilizes advanced 2nm and 3nm process nodes for next-generation AI infrastructure performance.

Context

AMD has officially unveiled its Helios rack-scale architecture, a next-generation AI infrastructure designed to challenge NVIDIA's upcoming Rubin platform. Leveraging cutting-edge 2nm and 3nm process nodes, the system integrates the Instinct MI450 series GPUs and EPYC Venice CPUs. This move signals a strategic shift toward high-density, rack-level solutions to compete with NVIDIA’s total-system dominance in the data center. The Helios rack is engineered for massive scale, delivering up to 2.9 exaFLOPS of FP4 performance and featuring 31TB of HBM4 memory—roughly 50% more capacity than the projected Vera Rubin NVL144. By utilizing 72 accelerators per rack and adhering to open standards like UALink and Ultra Ethernet, AMD aims to provide a more flexible and cost-effective alternative for hyperscalers. Deployment is slated for 2026, aligning directly with the expected arrival of NVIDIA’s Rubin architecture. Partnering with companies like HPE, AMD is positioning itself as a primary contender in the gigawatt-scale data center market, focusing on the trillion-parameter model training and inference segment that currently drives the bulk of AI capital expenditure.

Related Companies

Nvidia
Nvidia
NVDA
US
AMD
AMD
AMD
US
Amazon
Amazon
AMZN
US