AMD states its 2026 rack-scale AI solution delivers four times the energy efficiency of its 2024 platform, projecting a trajectory toward 20X improvement by 2030.
AMD has effectively institutionalized its “X-by-year” efficiency targets into a recurring engineering roadmap. In 2025, the company unveiled its 20x2030 initiative, pledging to increase the energy efficiency of its rack-scale AI solutions twentyfold by the end of the decade. Initially, AMD projected that its 2026 rack-scale AI systems would deliver three times the energy efficiency of its 2024 machines. However, updated estimates published on Tuesday indicate that the latest solutions will actually achieve a fourfold improvement over the 2024 baseline.
AMD measures progress at the rack level rather than isolating individual CPUs or AI accelerators, a methodology that allows broader optimization across the entire system. The company’s performance-per-watt calculations account for advancements in compute performance, process technology, memory bandwidth, data movement, interconnects, software, and system-level co-design. According to AMD, three primary hardware characteristics dictate AI system performance: compute capability, memory bandwidth, and interconnect bandwidth. New process technologies and architectures enhance floating-point performance per watt, while advances in memory integration and high-speed interconnects expand the bandwidth available to processors. Together, AMD expects these developments to yield 20 times higher AI performance per watt in 2030 compared to its 2024 baseline.
Memory and interconnects are particularly critical, as modern AI workloads require the transfer of massive data volumes between accelerators and host systems. AMD notes that higher memory bandwidth, greater bandwidth density, improved bandwidth per watt, larger caches, and tighter memory-compute integration can significantly reduce wasted energy and boost overall efficiency. Similarly, faster scale-up interconnects enhance communication between GPUs, CPUs, and other components, further improving performance efficiency. Software optimizations also play a vital role; by maximizing throughput through code and architecture tuning, AMD reduces the power required to achieve target computational results.
It is important to note that the reported fourfold efficiency gain remains an estimate rather than a direct benchmark between two commercially available rack systems. AMD tracks progress by comparing annual representative rack configurations against the 2024 baseline using its proprietary performance-per-watt methodology. Furthermore, the 2026 calculation blends measurements from shipped products with modeled projections where final performance data was still pending. Consequently, these estimates do not yet incorporate AMD’s latest Instinct MI455X accelerators.
Should AMD meet its 2030 objectives, the company projects that approximately two next-generation AMD racks could deliver the same computational capacity as 570 racks based on the 2024 Instinct MI300X. This translates to either a 20-fold reduction in power consumption or a 20-fold increase in performance at equivalent power levels.
In parallel development, AMD has officially introduced the Instinct MI455X AI accelerator and its accompanying Helios MI455X AI platform, which leverages UALink-over-Ethernet interconnects. Microsoft has also announced plans to deploy the Helios rack-scale AI accelerator at scale across its Azure cloud infrastructure.