O1.2.6 Demonstrate single (A100) GPU performant simulations with all parameterizations

[charleskawczynski]

The Climate Modeling Alliance

Software Design Issue 📜

Purpose

Demonstrate 0.5 s per timestep at ~110 km resolution with 64 levels with all (but gravity wave) parameterizations on a single (A100) GPU.

The approach will be to tackle both egregiously expensive gpu kernels, improve parallelism where needed, and use kernel fusion to reduce sequential reads/writes.

Cost/Benefits/Risks

People and Personnel

• Lead: @charleskawczynski
• Collaborators: @sriharshakandala
• Reviewers: --> @Julians42, @dennisYatunin, @nefrathenrici, @Sbozzolo

Components

Components considered include

• [ ] Moist dycore
• [ ] Prognostic implicit EDMF
• [ ] Radiation
• [ ] Diagnostics (if time permits)

[juliasloan25]

Is this a duplicate of https://github.com/CliMA/ClimaAtmos.jl/issues/2943 ?