>
The CEO of chip giant AMD has said the high-performance computing (HPC) industry needs to become significantly more efficient or it will have to turn to nuclear power.
AMD CPU and GPU performance typically doubles every 2.4 years, which is good progress by all means, but power efficiency hasn’t seen such significant improvements yet.
Dr. Lisa Su, speaking at the International Solid-State Circuits Conference, noted that with the zettaflop HPCs on the horizon, we need to turn our attention to the performance per unit of energy to achieve the zettaflop computation.
Go nuclear
Current trends and assuming progress doesn’t slow, AMD thinks the zettaflop compute could be here in about a decade (although Intel thinks it could be here in about half that time).
The reality is that when they arrive isn’t as important as when they become available on a usable, energy-efficient scale.
Currently, Su generously estimates that an HPC capable of performance measurable in zettaflops would use 500 MW of power, or the equivalent of several thousand homes. To deliver such a power consistently and reliably would probably require a nuclear reactor.
While there have been noticeable improvements in efficiency over the past few decades, progress has begun to slow down as the scope for further improvements shrinks.
“This leveling off of efficiency is going to be the biggest challenge we have to solve, both from a technology point of view and from a sustainability point of view,” said Su.
She explained that computing efficiency should become the number one priority to solve the problem. While our existing resources have slowed down, Su explained that there could be other solutions, including using artificial intelligence and machine learning for high-performance computing.
While there could be a nominal sacrifice to accuracy, working on improving this could be the way forward. Whatever combination of methods AMD and other companies ultimately use to deliver performance improvements, it’s clear that the journey between the first zettaflop-capable HPC and real-world use is likely to be a long one.
Through The register (opens in new tab)