Closed dirypan closed 3 weeks ago
You should be able to put them in the ***SolveArgs argument of ABM.
https://dsb-lab.github.io/CellBasedModels.jl/stable/API.html#CellBasedModels.ABM
Anyway, as far I understand, it is normal for GPUs to have such errors as they do not have the same error correction methods as CPUs. In Quadro ones, they have better error proof implemented.
Would this worry you to say the simulation results might be unreliable? My simulation is too costly on the CPU.
Sorry @dirypan, I would not worry much, it will always depend on the accuracy you need.
Normal GPUs work with float32, while CPU on float64. If you need the most extreme precission, then I will go to GPU-Quadro or try to go to CPU parallelizations in a server/supercomputer.
Hi, I found that with some integrators (say DifferentialEquations.Tsit5() ), changing dt will lead to different results, while the problem is not seen with the CPU platform. Here is the code that should reproduce the problem.
The above code will output
275.75217f0
for CPU platform no matter whatdt
is used (changesteps
andsaveEach
accordingly) and will output267.54324f0
for GPU platform withdt = 0.02
and output275.75208f0
whendt = 0.005
.Since both the CPU and GPU are using Float32, and this problem is not stiff. I wonder what could be the root of the problem. (I tried on A100 and 3090ti, and the same problem exists). If it is related to the default parameters of the integrators (say abstol/reltol), how can I modify them in the algorithm assignment?