AMD inches closer to its goal of making AI suck less … energy
AMD has announced significant progress in reducing energy consumption for artificial intelligence workloads, claiming its latest hardware generation achieves roughly four times the efficiency of systems from two years ago. The company launched Helios, a rack-scale platform containing 72 MI455X graphics processors, as part of an ambitious effort to improve overall datacenter efficiency by 20-fold before 2030. Whilst individual processors consume substantially more power, AMD argues that the dramatic improvement in how efficiently work scales across dozens of processors generates the overall efficiency gains.
However, important caveats accompany AMD's claims. The company employs a different methodology than rival Nvidia for calculating efficiency, weighting performance metrics differently rather than measuring actual application results. AMD itself characterises the 4x efficiency improvement as an estimate, with formal verification awaiting benchmark tests scheduled to arrive shortly after the systems begin shipping this quarter. These first independent measurements will determine whether AMD's ambitious efficiency targets prove realistic or require downward revision.
- AMD reports achieving 4x greater energy efficiency in its 2026 AI systems compared to 2024, working toward a 20x improvement target by decade's end.
- The company's new Helios rack-scale platform integrates 72 GPUs but uses different measurement methods than competitors; efficiency gains remain unverified pending benchmark results.
- Two Helios racks could theoretically replace 570 racks from 2024, though AMD's claims are preliminary estimates awaiting real-world performance validation.