I believe there's a hint towards the end of the article:
> Note: Perlmutter’s “AI performance” is based on Nvidia’s half-precision numerical format (FP16 Tensor Core) with Nvidia’s sparsity feature enabled.
FP16 is a 16 bit floating point format. FLOPS for top 500 are measured with LINPACK HPL, which says it is over 64 bit floating point values (I think):
> HPL is a software package that solves a (random) dense linear system in double precision (64 bits) arithmetic on distributed-memory computers. It can thus be regarded as a portable as well as freely available implementation of the High Performance Computing Linpack Benchmark.
I'm another user of the predecessor to Conducto. We're using it to orchestrate very large regression test (usually ~40-50k independent processes run as part of the regression testing).
I also use conducto's predecessor to manage "data science" workflows (simlar to https://www.conducto.com/docs/advanced/data-science) on a large computer. It is miles ahead of the alternatives we've tried (and built).
The biggest win we've gotten with conducto comes from the composability. We can ship part of our application as a function which returns conducto nodes. Someone else can call this function an stick it into their own tree (code, not configuration!)
Our integration test suite builds a conducto tree ships containing test code from 4ish different libraries (developed by 2 different teams), each of which includes it's own tree-building functions. Likewise, we ship many of our applications for the large computer as functions which return conducto nodes. Users just call our functions and embed our portion of the workload into their own trees.
CERN uses PTP on steroids for physics experiments: https://white-rabbit.web.cern.ch/