TOP500 at ISC’26: We have a New Number 1
10 points by johnklos
10 points by johnklos
The most significant part is that this system is also first place in regard to the HPCG benchmark, a much more representable benchmark than the usual Linpack crunching.
Let's see how the US responds.
To answer the author's question why US AI systems are not listed: Linpack and HPCG use FP64, and AI accelerators are very slow at such high precisions. You could use an Ozaki scheme, but on the other hand, these benchmarks are tightly integrated.
I am surprised that the new Italian system they mention runs on such old Software (RHEL 9).
Confused by the characterization of RHEL 9 as old, especially in the sciences. Its first stable release was in 2022 and importantly, is the current de-facto target for a huge amount of science software, even more so than RHEL 10. I work on a large instrument and we only just moved most of our infrastructure over from 7 to 9. ("most" is doing so much work. We still have several important systems on Scientific Linux 6).
I work in a big enterprise and we are also on RHEL9 but I was always assuming that HPC folks would chase the latest and greatest compilers and kernel features to squeeze more out of the hardware. Maybe my assumptions are completely off.
In HPC, the system compiler is rarely used; new compilers are typically provided as modules or built via Spack, in addition to the compilers provided by the vendor. Kernel features.. HPC is one of the domains that likes bypassing the kernel where possible. The OS really doesn't matter too much in the end, it just might be a minor annoyance for users.
The memory access times seem glacial compared to modern hbm4/whatever memory. On the top model from China.