I have Ge 8800GS - 96 cores arranged in 12 MP.
Now I ran CUDA SDK device Query and it tells me that maximum threads per block = 512
and that max dimension of a block is 512 * 512 * 64.
Now if max dimension of a block is 512 * 512 * 64 => no. of threads should be this much rather than 512???
Where am I wrong?Help please.
from spec of compute capability 1.0, you have
The maximum number of threads per block is 512
The maximum sizes of the x-, y-, and z-dimension of a thread block are 512, 512, and 64, respectively
these two conditions (constraints) must be satisfied at the same time.
for example:
legal: 16 x 16 = 256, 4 x 4 x 8 = 128
illegal: 32 x 32 = 1024 (condition 2 fails), 8 x 8 x 15 = 1024 (condition 2 fails)
the maximum number of threads per block comes from hardware limitation since
GPU must provide synchronization operation for threads of one block, so this number cannot be so large,
512 is O.K. since it need 9-level AND gate to make sure synchronization
Thank you LSChien. So the deviceQuery function is creating confusion when it says max sizes of each dimension of a block is 512 * 512 * 64. It should have been 512 for dimension x , same for y and 64 for dimension z.
yes, it indeed said what you think about
The maximum sizes of the x-, y-, and z-dimension of a thread block are 512, 512, and 64, “respectively”
in fact, some conditions in spec of compute capability must be considered together,
I can show you another
from spec of compute capability 1.0
The maximum number of active blocks per multiprocessor is 8
The maximum number of active warps per multiprocessor is 24
The maximum number of active threads per multiprocessor is 768
if you want to ask me “what is maximum number of active blocks in a multiprocessors”, then
you must combine these three conditions. however condition 2 is redudant, because from
condition 3 and 32 threads/warp, you have 768/32 = 24 warps
hence we only consider condition 1 and condition 3
example 1: suppose we choose size of threads block as 16 x 16 = 256, then
under condition 3, we have 768/256 = 3 blocks in a multiprocessor
under condition 1, 3 < 8, hence wew have 3 blocks in a multiprocessor
example 2: suppose we choose size of threads block as 4 x 4 = 16, then
under condition 3, we have 768/16 = 48 blocks in a multiprocessor
under condition 1, 48 > 8, hence wew only have 8 blocks in a multiprocessor
this means that we only have 8 * 16 = 128 active threads in a multiprocessor
Can I define a block size like dim3 dimBlock(20,20,1)? because I have been trying this, and do nothing, if I define the dimBlock dim3 dimBlock(20,12,1) works fine, but, after dim3 dimBlock(20,13,1) stops working, is like if the maximum number of threads per block were 256
you can do max of max
square_array <<< dim3(4096,16,1),dim3(16,32,1) >>> (memoiregraphique1, N,in,ss);
square_array <<< dim3(1024,64,1),dim3(64,8,1) >>> (memoiregraphique1, N,in,ss);
=40961616*32 thread
409616=65536 = 102464 =65536 max
1632=648=512 max for me perhaps 256 for you
When you submit the kernel you use a parameter dim3 (tx,ty,tz) which specifies the number of threads in a block. The device query only says that tx<=512, ty<=512 and tz<=64 and in addition to this 3 you also must have txtytz<=512.
So you have 4 conditions for the threads in a block:
tx<=512
ty<=512
tz<=64
and
txtytz<=512.