Forum > GPU crunching
GPU client
Gecko_R7:
Mimo,
In what order does clock speed impact GPU performance as far as S@H is concerned? CPU clock, memory clock, shader clock?
Also, do I understand correctly that the G92 8800GT has 12 FPU processors in the GPU?
Do the shaders provide any benefit?
Sorry for the questions. Just tying to understand this better.
Devaster:
yes . for example 8500 GPU have two multiprocessors where every multiprocessor has 8 unified shaders, every shader can work with 4 floats at one instruction - you may imagine that your cpu has 16*4 cores ....
instructions are very effective - low clocks time, for example MADD - multiply and add have only 4 clocks.
cache is extremely effective in contignous reads/writes - called coalescing
clocks speed havent so great impact on gpu performance as count of shaders
Gecko_R7:
--- Quote from: Devaster on 18 Dec 2007, 02:12:23 pm ---yes . for example 8500 GPU have two multiprocessors where every multiprocessor has 8 unified shaders, every shader can work with 4 floats at one instruction - you may imagine that your cpu has 16*4 cores ....
instructions are very effective - low clocks time, for example MADD - multiply and add have only 4 clocks.
cache is extremely effective in contignous reads/writes - called coalescing
clocks speed havent so great impact on gpu performance as count of shaders
--- End quote ---
Thanks Mimo! Very interesting. So, higher shader count and faster shader clock will actually have better impact on crunching speed/potential for our purposes? In the case of a new G92 based 8800GT, 112 stream processors, each that can process 4 floats in 1 instruction. Wow! The interest in this becomes very clear. G80/G92 stream processors are scaler units, not vector processors?
Devaster:
more about theory is in this document on the beginning : http://developer.download.nvidia.com/compute/cuda/1_1/NVIDIA_CUDA_Programming_Guide_1.1.pdf
no shaders are strictly vectorized , massive pararelizm is implemented on hw....
Gecko_R7:
--- Quote from: Devaster on 18 Dec 2007, 05:34:50 pm ---more about theory is in this document on the beginning : http://developer.download.nvidia.com/compute/cuda/1_1/NVIDIA_CUDA_Programming_Guide_1.1.pdf
no shaders are strictly vectorized , massive pararelizm is implemented on hw....
--- End quote ---
Thanks!
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version