Simulation cost estimator

This interactive tool allows to quickly estimate the cost of a simulation (both in terms of memory usage and simulation time) to help optimize your setup for specific scales and architectures.

Basic parameters

Mode:

Dimension:

Coordinates:

$r_0$: 

$h$: 

Grid & particles

Resolution:

× ×

CFL:

Extent in $x^1$:

Total runtime:

# of particles per cell:

# of payloads:

Precision:

single real_t

single prtldx_t

Architecture

GPU arch:

V100 [32GB] A100 [40GB] A100 [80GB] H100 [80GB] H100 [96GB] H100 [160GB] MI250X [128GB] Ponte-Vechio [128GB] custom

GPUs per node:

Number of GPUs:

Custom parameters:

per GPU memory:

GB

timestep per particle:

ns

memory overhead:

%

performance overhead:

$\times$

Memory usage

Per cell/particle

Fields: ...

Particles: ...

Total

Fields: ... ...

Particles: ... ...

Total: ...

Per GPU: ... ...

Runtime

Timestep duration

$\Delta t$: ...

Duration: ...

Total

Recommended number of GPUs: ... ...

Number of timesteps: ...

Runtime: ...

Cost: ...

Memory alignment

Because of architecture-dependent memory alignment, the numbers cited here are only approximations. The actual memory used by the run may vary, so please make sure to leave at least 20%...30% of headroom for safety.

Accuracy

Current performance estimates for specific architectures are very crude, and we will most likely refine them in the future. Additionaly, keep in mind that the efficiency is proportional to the number of particles you can fit on a given GPU, so make sure to maximize the usage as much as possible for the best performance.