AMD's AI systems are already 4x more energy-efficient than 2024 baseline

Full text

AMD says its rack-scale AI systems are now about four times more energy-efficient than its 2024 baseline, beating its own threefold interim target by 33%

The company offers two alternative outcomes for that gain: the same compute in far fewer racks, or twenty times more compute for the same power, something that

AMD's own figures can also be reverse-engineered to imply that each 2030 rack draws roughly 14.25x the power of a 2024 one

AMD has published a progress report on the efficiency goals it set in June 2025, revealing the numbers are already better than the company was aiming for.

Against a 2024 baseline, AMD says its rack-scale AI systems are now roughly four times more energy efficient, ahead of the three times it had projected for this point, a trend that, by its own calculations, represents more than double the historical industry trendline.

This aligns with a six-year commitment by the chip design giant, in which it has pledged to deliver a twenty-fold improvement in rack-scale energy efficiency for AI training and inference by 2030.

Efficiency gets center stage as AMD aims to secure more AI customers

This is not the first time AMD has set itself a performance and efficiency goal. Its previous target, internally known as 30x25, aimed at a thirtyfold node-level improvement between 2020 and 2025, and the chipmaker finished ahead of its projections, clocking in at 38x ,

AMD's efficiency and performance gains, if they hold as intended at 20x, can be flipped two ways: one could consider sizing down to larger, but singular racks to get their job done versus legacy hardware.

The flipside is easy to make at a time when everyone from AI startups to hyperscalers is pushing to increase compute whenever possible. AMD's efficiency gains could, in principle, allow its latest offerings in 2030 to deliver 20x the compute for the same power draw.

At a time when nobody in the industry is trying to buy less compute, AMD's offerings might be particularly appealing thanks to their efficiency gains, as projections indicate global data center electricity demand will more than double by 2030 to around 945 terawatt-hours, roughly what Japan uses today as a nation.

AMD's gains come from its newest GPUs being an order of magnitude more efficient than their predecessors; the MI300X that anchors the baseline was a 750-watt part delivering up to 2.6 petaFLOPS of dense FP8.

For context, each current-generation MI455X in Helios offers between 7.7 and 15.4 times the floating-point performance, 2.25 times the HBM, 4.4 times the memory bandwidth, and four times the chip-to-chip interconnect bandwidth, while consuming three times the power.

Its chief sustainability officer, Justin Murrill, told Trellis that every AMD team carries efficiency-per-watt targets for new products and that progress is tied to company-wide bonuses.

For now, AMD is ahead of schedule on a real, well-documented engineering goal. Hitting it, however, will not reduce the amount of electricity the AI industry consumes, as the efficiency and performance gains it offers will only fuel more demand for AI hardware that can use the saved power, even as demand for AI compute power continues to soar.