2 km UK: what it would buy, and what it costs

4 km is the workhorse resolution for Convek. It is the point where a regional model starts producing the mesoscale features a global model cannot - sea breezes, convergence lines, valley flows - at a compute cost that fits four cycles a day on a single worker. Everything live today, six countries' worth, runs at 4 km. The obvious next rung is 2 km, and it is worth being clear about what that actually buys and what it costs before treating it as a done deal.

Start with the cost, because it is the part that decides everything else. Halving the grid spacing makes a WRF run roughly eight times slower: four times the columns, and a time step that has to halve to stay numerically stable. The live trimmed UK 4 km cycle lands in around 34 minutes end-to-end on the `raspuk` worker. A 2 km version of the same domain is estimated at four to five hours per cycle on a bigger box - and that estimate is a planning number from the compute scaling, not something measured on the live hardware yet. Four to five hours does not fit a four-cycle-a-day schedule on the current worker, full stop.

So 2 km is not a namelist change, it is an infrastructure decision. It needs a second, heavier worker - the plan is a VPS6-class box with more vCPU and memory - dedicated to the high-resolution domain, running alongside the existing 4 km batch rather than competing with it for the cycle window. And it needs `gb/4km` to be genuinely rock-solid first, because there is no sense building a more expensive product on top of a base layer that still occasionally needs hand-holding. The order is: stabilise 4 km, stand up the second worker, benchmark 2 km for real, then decide.

What it would buy pilots is terrain. At 4 km, a grid cell is about 16 km² - good enough for the sea breeze and the broad convergence line, but it still smears the things that happen at the scale of a single ridge. Ridge convergence zones, the trigger points on a south-facing face that fire an hour before the valley, the sharpness of a sea-breeze front rather than just its presence, valley wind detail in the Welsh and Lake District terrain. 2 km starts to resolve those. Cloudbase also tightens up, because the moisture profile is sampled on a finer grid where the terrain forcing is sharper.

On horizon, the 4 km product now targets a 72 hour window. The 2 km question is whether a new box can hold that same roughly three-day window without crowding the 4 km production batch. That is a runtime question, not a physics one, and it is downstream of the benchmark that has not been run.

The honest status, then: 2 km UK is planned, not in development, and not benchmarked on the live infrastructure. The numbers in this post are scaling estimates. When the second worker exists and the first real 2 km cycle runs, this post gets rewritten with measured wall-clock and a side-by-side against the 4 km output - because the only thing that justifies eight times the compute is evidence that the finer grid is actually a better forecast, not just a prettier one. When it ships it will be a Pro-tier resolution alongside the free 4 km base.

The discipline here is the same one that runs through the rest of this blog: ship the cheap thing solid, validate it, and only buy the expensive upgrade once there is a number to chase. 2 km is the upgrade. The validation pipeline and a stable 4 km base are the prerequisites. The optimisation series covers the rest of the stack, and the model page has the current shipping configuration.