Top500? Maybe Mid500

We’re days away from the start of SC25 which means we’re days away from the release of a new Top500 list too. While I like seeing that as much as the next HPC Nerd (Patent Pending: Boof) what does it really mean anymore?

It’s been about six years since China stopped submitting their fastest systems. A look at the June 2025 list shows almost all the entries are from governmental organisations with the few exceptions being technology companies that are using it to showcase their wares. Even then, Microsoft’s Eagle supercomputer which was based on their NDv5 cloud offering represented only a fraction of the total size.

The most obvious missing candidates are the AI behemoths such as Colossus xAI or Microsoft Fairwater. Whilst they may not be as big, there are still a number of systems in the private sector (especially in banks and hedge funds) that would comfortably be in the top 20.

Part of the problem here I think is that the HPC community often feels a bit like an academic science clique with all that entails about people not wanting to be a part of it and setting up their own little cliques instead. Linpack (the benchmark used for the Top500) may have some relevance in academia but AI training and financial risk really couldn’t care less about it. This not only means that the system may not be optimally tuned for Linpack performance but also that the people responsible for it will be less inclined to bother running it.

If supercomputing is going to be about more than just academia and continue to exist as a field rather than be subsumed in various parts into cloud, AI, devops, SRE and parallel programming then maybe we need to rethink using a benchmark from the 1970s.

I’m aware we also have HPCG but the candidates on that list are essentially the same as those on the Top500. There’s no incentive or draw for those outside of the current HPC community to want to participate it seems.

Of course, many in the private sectors won’t want to publicise their capabilities regardless of how relevant they may be. Perhaps historic submissions might be useful, submit results for your last gen system as you build out the new one? Or partial runs like that done for Eagle I guess.

I might be eating my words in a few days if the next Top500 list comes out with xAI, Fairwater and some quant hedge fund topping the list! Actually, I hope that does happen even if it makes me wrong.