Re: [ESPResSo-users] CUDA error: out of memory

[Stefan Kesselheim]

Hi,

On Aug 2, 2013, at 11:39 AM, Markus Gusenbauer <address@hidden> wrote:

> You are absolutely right, it is a 2GB GPU, thanks for the info! Can I use 
> more than 1 GPU for a simulation?

Yes, and that is what you are doing. You get the error because your GPU memory 
is full! If you need more than 200^3 you need a GPU with more memory as we 
cannot do multi-gpu (yet).
Cheers
Stefan

> 
> Markus
> 
> 
> On 02.08.2013 11:17, Stefan Kesselheim wrote:
>> Hi Markus,
>> yes, that is quite right. We need 19xN single precision float numbers per LB 
>> cell, times two, because we stream from one lattice to a second lattice to 
>> avoid race conditions.
>> 200**3 cells *19 populations * 4 byte * 2 buffers=1216MB,
>> We need a bit more to save forces and so I'd conclude you have a 2GB GPU.
>> Cheers
>> Stefan
>> 
>> On Aug 2, 2013, at 11:01 AM, Markus Gusenbauer <address@hidden> wrote:
>> 
>>> Hi all,
>>> 
>>> is the number of nodes with the LBM on GPU limited by hardware? When I 
>>> choose too large simulation size, I get a CUDA out of memory error.
>>> Right now I am limited to around 8,320,000 nodes (200x200x208), after that 
>>> it crashes.
>>> 
>>> Example:
>>> 
>>> setmd time_step 0.1
>>> setmd skin 0.2
>>> thermostat off
>>> 
>>> set boxX 200
>>> set boxY 200
>>> set boxZ 208
>>> 
>>> setmd box_l $boxX $boxY $boxZ
>>> 
>>> lbfluid gpu grid 1 dens 1.0 visc 1.5 tau 0.1 friction 0.5
>>> 
>>> set cycle 0
>>> while { $cycle<100 } {
>>>    puts "$cycle / 100 \r";
>>> 
>>>    for { set i 1 } { $i < [expr $boxX-1]} { incr i } {
>>>        for { set j 1 } { $j < [expr $boxY-1] } { incr j } {
>>>            for { set k 0 } { $k < 1 } { incr k } {
>>>                lbnode $i $j $k set u 0.0 0.0 0.001
>>>            }
>>>        }
>>>    }
>>> 
>>>    integrate 1
>>>    incr cycle
>>> }
>>> 
>>> 
>>> Thanks!
>>> 
>>> Markus
>>>