Relationship between sp, warp, grid tell the difference.

[fzhsheng]

My video card is Geforce 8800GT.

I read the programer’s guide:

but I donnot understand the difference between Stream Processor,

warp, and Thread Processor:

I know 8800GT has 112 stream processors;

Run the deviceQuency in sdk demo, It tells me there’s 32 warps. what’s the

relationship between warps and grid?

and I donnot know how many Thread Processors it have.

What’s the relationship between them?

[DenisR]

warp size = 32 → 32 threads run ‘at once’ (actually 8 run at once, and 4 cycles after eachother are the same instruction)

in cuda you have blocks of threads. A block can have at most 512 threads.
You can have 32768 * 32768 blocks

Then you have multiprocessors (with 8 processors on each). A block runs on a single multiprocessor, but there can be more than 1 block running on a single multiprocessor.

A GPU contains up to 16 multiprocessors (8*16 = 128 processors) Yours has 14 multiprocessors.

[fzhsheng]

warp size = 32 → 32 threads run ‘at once’ (actually 8 run at once, and 4 cycles after eachother are the same instruction)

in cuda you have blocks of threads. A block can have at most 512 threads.

You can have 32768 * 32768 blocks

Then you have multiprocessors (with 8 processors on each). A block runs on a single multiprocessor, but there can be more than 1 block running on a single multiprocessor.

A GPU contains up to 16 multiprocessors (8*16 = 128 processors) Yours has 14 multiprocessors.

[snapback]374408[/snapback]

I see. Thank you…