AMD reports fourfold efficiency boost in AI rack systems since 2024
AMD says its latest rack-scale AI platform is about four times more efficient than its 2024 baseline.
AMD announced that its newest rack-scale AI compute platform, codenamed Helios, delivers an estimated fourfold efficiency improvement over the 2024 baseline, according to the House of Zen blog. The Helios system integrates 72 MI455X GPUs, each offering between 7.7x and 15.4x higher floating-point performance, larger HBM capacity, faster memory and quadruple interconnect bandwidth compared with the earlier MI300X. While individual GPUs consume more than three times the power, the overall system gains stem from more effective scaling of AI workloads across all accelerators.
AMD attributes these advances to hardware upgrades, new 4-bit floating-point support, memory technology enhancements, and a refined software stack. The company’s broader ambition is to reach a 20-fold increase in rack-scale efficiency by the decade’s end, positioning it against competitors such as Nvidia’s Blackwell-based NVL72 racks.
Why it matters
Improved AI hardware efficiency can lower data-center power costs and reduce environmental impact.
In this story
