April 17, 2026
Million cores, same old code?
30 Years of HPC: many hardware advances, little adoption of new languages
Supercomputers got huge; coders stick to C & Python—and the comments are on fire
TLDR: A Chapel blog retrospection says supercomputers exploded in power while most programmers still use the same old languages. Commenters clap back that hiring is easier for C/C++ and Python, memory bandwidth—not new syntax—rules performance, and many real labs run scripts over shiny new languages.
Chapel’s Brad Chamberlain looked back at 30 years of high‑performance computing and went full “we can put a man on the Moon but still write in C.” The hardware glow‑up is wild—millions of cores, jaw‑dropping speed—but the big reveal is that new languages haven’t won hearts. Cue the comments section riot.
The top hot take? “New languages don’t fix real pain.” One veteran insisted the bottleneck isn’t clever syntax but memory bandwidth, dropping the mic with “the code is often boringly single‑threaded because that’s fastest.” Another thread caught fire over hiring: why bet on niche tools when you can find a C/C++ dev who can learn OpenMP in a week? A crowd favorite meme: HPC = Highly Persistent C.
Then came the reality check: on many clusters, it’s Python all the way down—with R, Perl, and even awk cameos. One bioinformatics commenter admitted “MPI barely reached us” as researchers prefer one giant machine to “proper” distributed wizardry. Meanwhile, another wondered why actor‑style ideas (think Erlang) never took off for millions of cores—an entire debate spawned about message‑passing dreams vs. bandwidth‑bound reality.
An early‑’90s editor chimed in with nostalgia, but the mood stayed spicy: million cores, same old code. New languages brought cool logos; the old guard brought receipts.
Key Points
- •The article reflects on 30 years of HPC programming, drawing on a HIPS 2025 keynote and a CLSAC 2025 lightning talk.
- •It compares TOP500 top five systems from November 1995 and November 2025 to illustrate hardware evolution.
- •1995 top vendors included Fujitsu, Intel, and Cray, using crossbar, 2D mesh, and 3D torus interconnects, with 80–3,680 cores and 98.9–170 GFlop/s Rmax.
- •2025 top vendors included HPE Cray, Eviden/Bull, and Microsoft, using Slingshot‑11 and InfiniBand NDR with dragonfly and/or fat‑tree topologies, with 2,073,600–11,340,000 cores and 561–1809 PFlop/s Rmax.
- •Over 30 years, core counts increased by roughly 563–141,750× and Rmax performance by about 3.3–18.3 million×.