Next-generation computing system
Next-generation computing system
review three to four upcoming supercomputer designs. Archer 2 in the UK, Perlmutter and others show different approaches—almost identical chips with minor tweaks. They typically feature around 55 to 75 cores, which is why Epic achieves such speed. It's not a single-core design and operates at 50% higher clock speeds.
they used to all run AMD models from 10 to 15 years back. they switch to the latest stable platform that performs best. they don’t adopt the newest chip right away; instead, they move to a second or third generation version quickly. they rarely upgrade to the very newest chips but focus on those with proven stability. their Xeon Phi received its final three updates canceled because EPYC dominated and many supercomputing bids won.
They adapt to whoever appears most reliable at any given moment. Fifteen years ago AMD avoided major mistakes. However, with FX they made errors. Now they must rebuild trust. This isn't about raw speed. In the coming years, supercomputers will rely on EPYC. But that's not the main focus. True advancement lies in stability and understanding your system's capabilities. Optimizing software can yield remarkable outcomes, even against newer hardware. As a company investing millions or billions in next-gen supercomputers, you can't simply accept untested products like EPYC. Consider the stakes—this isn't about small amounts. Remember, most supercomputers serve military purposes, and any shortcut could lead to serious consequences. EPYC is already being adopted by them. The more AMD demonstrates its worth, the greater the confidence they'll gain. Give it time.
It wasn't about reliability. I visited Berkely Labs' supercomputer and inquired if, according to their roadmap, Epyc would become the next major chip in most supercomputers. Two years prior to its announcement, I was confident it would be the upcoming choice. Spoiler alert!