Within six years, AMD rack solutions will be 20 times more efficient

In addition to its consistent product and technology roadmap, AMD also launched the “20×2030” plan in 2025, promising to increase the energy efficiency of rack-mounted artificial intelligence solutions by 20 times by 2030. AMD expects its 2026 rack-mounted systems to be three times more powerful than the 2024 generation. In practice, they are four times more efficient, allowing us to count on plans to overtake AMD by 2030.

Image source: AMD

It’s worth mentioning that AMD evaluates its progress at the rack level rather than at the processor and AI accelerator level, which obviously gives the company more freedom in optimizing performance. The company is considering improvements in computing performance, process technology, memory bandwidth, data movement, interconnect, software and system-level co-design.

AMD said that three key hardware characteristics determine the performance of AI systems: computing power, memory bandwidth and interconnect bandwidth. New process technologies and architectures are increasing floating-point performance per watt, while improvements in memory consolidation and high-speed interconnects are increasing the bandwidth available to chips. AMD expects that by 2030, these developments will achieve a 20x improvement in AI performance per watt compared to the 2024 baseline.

Memory and interconnect play a particularly important role because modern AI systems tend to move large amounts of data. AMD said higher memory bandwidth, greater bandwidth density, improved bandwidth per watt, larger caches and tighter integration of memory and computing resources can reduce power waste and improve operating efficiency.

Faster, scalable interconnects improve communication between accelerators, CPUs, and other components, resulting in greater operational efficiency. Software optimization AMD is another part of the effort, as higher performance through optimization ultimately means less power required to achieve the desired results.

The four-fold improvement in energy efficiency (and thus performance) should be viewed as an estimate from AMD rather than a direct comparison of two commercially available rack-mounted systems. AMD measures progress by comparing annual representative rack configurations to a 2024 baseline using a performance-per-watt methodology.

Additionally, without final performance data being available, the 2026 estimates used actual product measurements and simulation results, which essentially means AMD didn’t use its latest Instinct MI455X accelerator to make these estimates.

If the 20×2030 plan is successful, AMD’s two future 2030 racks could provide the same computing capacity as the 570-processor-based rack in 2024 Instinct MI300X. This essentially means that by 2030, for the same energy consumption, energy consumption will be reduced by a factor of 20, or efficiency improved by a factor of 20.