I am looking for an order of magnitude estimate for the cost or running a scale resolved DNS solution on an object with transition and significant separation at a Re of 400k. For units I am thinking perhaps A100 hours may be a good choice. For example, how many hours would it take to run a sphere at Re=400,000 on (32) A100’s? 1 hr? 10 hrs? 100 hrs? 1,000 hrs? 10,000 hrs? I would want enough time steps to get a reasonable estimate of the average drag coefficient.
Baseline resolution requirements can be had here which concerns flow over a sphere. These simulations were mostly performed on 32 V100 GPUs (slower than an A100) and while the precise amount of time varied between set ups I recall them taking about a week on average.
Regards, Freddie.
Freddie,
This was very helpful. I have a lot of experience testing spheres with a very accurate system (not a wind tunnel). The subcritical CD in the paper is within 1% of what I quote based on experimental results.
Thank you,
Steve