Single precision

[jagoosw]

Hi all,

I was wondering if anyone has done any validation in single precision? The last I heard no one had validated if the truncation is a problem.

I have access to some Nvidia L40S which are faster than A100s for single precision but can't do double precision and was thinking about trying to run some of my models in single precision instead.

Thanks!

[glwagner]

It depends on the simulation you want to run, and there are a wide variety you may consider with Oceananigans.

We did a bit of investigation on https://github.com/PRONTOLab/GB-25 and found that single precision was unstable for a hydrostatic model using WENO-only dissipation. 9th order was quite unstable, but some cases with purely 5th order WENOVectorInvariant also exhibited more instability than with Float64.

I think it's possible that 3D turbulence cases with NonhydrostaticModel are more stable than quasi-2D cases with the hydrostatic model + WENOVectorInvariant. But you may have to adjust the WENO order and you may find that the solutions are more dissipative.

I'd suspect that DNS is less sensitive, but I'm not sure. The reason being that solutions are smooth at the grid scale and we probably want second-order advection, not WENO.

One key question is whether the likely increased level of numerical dissipation one gets with single precision (and therefore the need for higher resolution) offsets the savings of single precision, or whether single precision still provides some speed up.

[jagoosw]

That's interesting, thank you. I have some cleaning up to do to make sure that Walrus is type-stable, but then I'll try running a nonhydststatic model (Centered + AMD so less numerical dissipation?) in single precision and report back.

[glwagner]

(Centered + AMD so less numerical dissipation?)

Note that Centered + AMD has more total dissipation (numerical + explicit dissipation).

But I am not sure the issue regards numerical versus explicit dissipation, but the fact that you may need more total dissipation to stabilize a single precision scheme. WENO(order=9) has less total dissipation than Centered + AMD (which is why it has 2x higher effective resolution), but this also may be the reason why it is less stable.

[glwagner]

Be cautious about Centered + AMD for cases that are sensitive to dissipation by the way. Pressel et al 2017 have some cautionary tales: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2016MS000778

As well as Silvestri et al 2024: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023MS004130