Geek asks and answers

GPU Architecture (Nvidia)

In all the documents that I read, I see that the GPU consists of multiprocessors, and each multiprocessor has 8 processors that are capable of performing a single warp in parallel.
The graphics processor I use, the Nvidia 560, has only 7 multiprocessors, but 48 processors in each multiprocessor. Does this mean that every multiprocessor in the Nvidia 560 is capable of 6 distortions in parallel?
Can I say that the maximum number of threads running in parallel on the Nvidia 560 is 32 * 6 * 7 = 1344 threads in parallel? (32 = warp, 7 = multipliers, 6 = warps executed in parallel)

How many multiprocessors are in the fastest Nvidia GPU? What is this GPU? What is the maximum global memory that the largest GPU has?

+3
source share
2 answers

The documents you are reading are old. The first two generations of CUDA GPUs had 8 cores per MP and wrote out instructions from one warp (if you want to simplify, each command will be executed four times on 8 cores to serve one warp).

Fermi. " " ( warp 16 ). , 16 , .. . 2.1. 2.0 32 . MP- - . , , , 48 .

, - 14 , 336 7 GTX 560, , " " . , , , F .

+2

CUDA 4.2:

[...] , warp warp , ( warp) .

, waprs SM warp (WS).

GeForce 580 2.1:

2.x : [...] 2

, SM GPU 2 warps = 64 , 448 . , , , :

, , warp , , warp warp , , , "".

: GeForce GTX690 3072 CUDA Cores. CUDA 1536 , GeForce 680, , -. ​​ nVidia Tesla M2090: 6GiB GDDR5 (512 CUDA Cores). , Teslas, , GeForce 6xx, , .

+3

More articles:


All Articles