If all goes to plan, AMD’s supercomputer and server chips will become 30 times more energy efficient by 2025, thanks to upcoming silicon improvements.
On Wednesday, AMD announced the ambitious goal for its Epyc CPUs and Instinct GPUs in the hopes it’ll help new supercomputers and data centers achieve high performance, without straining the world’s energy supplies. Specifically, AMD is aiming to hit energy efficiency benchmarks when its comes to chips running AI-training programs and high-performance computing applications.
“The 30x goal would save billions of kilowatt hours of electricity in 2025, reducing the power required for these systems to complete a single calculation by 97% over five years,” AMD says.
The goal arrives in the context of a chip’s performance per watt. AMD is striving to raise this benchmark by 30 times within the next five years. So the server chips will still use plenty of electricity, but can do so while exponentially boosting performance. (Conversely, data centers could use fewer servers at lower energy levels and still achieve the same performance.)
Chip makers are constantly making improvements to their products to raise the performance-per-watt ratio. However, AMD’s goal means it’ll need to outperform the rest of the industry on server energy efficiency at a rate that’s 2.5 times faster over the next five years.
The big question is how the company will pull this off. AMD itself acknowledges that it won’t be easy. “Even with continued advances in process manufacturing, the slowdown in Moore’s Law is clear. Energy-efficiency gains from process node advances are now smaller and less frequent,” the chip vendor says.
“Therefore, a larger fraction of improvements needs to come from silicon architecture and packaging innovations in addition to expected gains from silicon process technology,” it adds.
The company didn’t elaborate on the upcoming improvements. But over the years, AMD has filed patents for its “Exascale Heterogeneous Processor.” The technology promises to boost a supercomputer’s efficiency by packing the CPU, GPU, and memory on a single chip.
