✓ ShippedByAgents
Practitioner (own blog/paper) Language port

University of Tokyo researchers port the NICAM climate model from 140,000 lines of Fortran to 35,000 lines of C++/Kokkos with Claude Code, running 1.71x faster on GH200

Researchers at the University of Tokyo and Japan's National Institute for Environmental Studies used Claude Code, mainly with the Fable 5 model, to port a production configuration of NICAM, a legacy Fortran/OpenACC atmospheric model, to C++/Kokkos. The human set scope and constraints in a kickoff document, approved a milestone plan written by the agent in plan mode, and reviewed each module; the agent translated module by module, building and testing against reference dumps from the Fortran model. About 140,000 lines of Fortran were analysed and ported into about 35,000 lines of C++/Kokkos covering the dynamical core and seven physics packages. Two 100-member ensembles showed the port is statistically consistent with the original, and it ran 1.71x faster on NVIDIA GH200 and 1.54x on AMD MI300A than the Fortran baseline on GH200. Claude Code wrote all of the C++/Kokkos model code.

The numbers (as reported)

Code ported~140,000 lines of Fortran analysed and ported into ~35,000 lines of C++/Kokkos
ScopeDynamical core and seven physics packages
Application speedup vs Fortran/OpenACC on GH2001.71x on GH200, 1.54x on MI300A
Active workflow days23
Human instructions797 (6,007 assistant messages)
Token cost~USD 5,589 estimated at API list prices (used under a subscription plan)
Ensemble agreement95.3% of 5,900 member–quantity scores within |z| ≤ 2 in both implementations

Facts

Company / subject
University of Tokyo and NIES (NICAM)
Tools
Claude Code
Models
Fable 5
Codebase
brownfield
Work type
Language port
Stack
Fortran, OpenACC, C++, Kokkos, CUDA, HIP, MPI, YAML
Published
2026-10-30
Last checked
2026-11-16

What the source claims Practitioner (own blog/paper)

Claimed by the builder, writing about their own work. Not checked independently.

  • Port runs 1.71x faster on GH200 and 1.54x on MI300A than the Fortran/OpenACC baseline 1.71x / 1.54x
  • Port is statistically consistent with the original Fortran model in 100-member ensemble tests
  • One codebase now runs on both NVIDIA and AMD GPUs through Kokkos CUDA and HIP backends

Caveats

Written by the researchers who did the port. The cost is an estimate at API list prices; they actually used a subscription plan. Human working hours were not recorded (instruction counts come from Claude Code's own logs), and days spent only on human review or cluster work are not in the 23-day count. Branches outside the selected configuration were not ported, and the authors credit NICAM's consistent, well-maintained Fortran for part of the smooth run. Human review and the subscription's five-hour usage limits were the main constraints.

More stories by tool: Claude Code

Related topics: Language portsBrownfield codebases

Source

arXiv (Ren, Shimokawabe, Yashiro)

One email a week: what actually shipped

New sourced stories, the numbers behind them, and what didn't work. No hype.

We'll only use this to send the newsletter. See the privacy page.