NWP Turbulence granule Python port

pitch-4a7b90 · in MCH

Status: Thinking — at the foot of the hill Shaping — halfway up the hill Ready — at the top of the hill In progress — halfway down the hill Done — at the bottom, over the hill Shelved — off the hill Somebody is building this in the cycle it was bet into — an overrun is re-bet, not extended.
Priority: Medium
Appetite: 0.5 person-weeks A budget in person-weeks, not an estimate: what this is worth, not how long it takes.
Tags: icon4py, turbulence, nwp, granule, blueline, port, mch
Owner: muellch
Assignees: muellch
Reviewers: OngChia
Review waived: —
Parent: MCH
Blocked by: —
PRs: —
Cycle: 38 A bet is for one cycle. Work that overruns is re-bet, not extended by default.
Start date: 25.08.2026
End date: —
Scheduled: 2026-08-25 → 2026-08-31
Blocks: —
Progress: 0/1 items

Problem

The NWP 1D turbulence is badly optimized in OpenACC. Fusing the pointwise stencils could give a large speedup -- claude estimates 3x, and 1.4x would already be worth having.

Appetite

Half a week.

Solution

Instruct claude to port the code to icon4py and write a blue line granule interface. Port only living code: many of ICON's namelist switches are never used. Verify against serialized data from the MCH and DWD experiments.

Rabbit holes

Avoid extensive rewriting, reformulating and refactoring. Even where it is ugly, staying close to the Fortran is probably right for now.

No-gos

Interfacing to the green line driver. Not yet.

Progress

No PR yet

This row sits below the cycle-38 betting table rather than in it, with no developer column filled, so it was not formally bet. It is written here because Christoph is in fact working on it.