
How to calculate theoretically possible GFLOPS
bpurnomo Apr 21, 2010 9:07 PM (in response to noxnet)For HD5850, there are 1440 Streaming Processing Units (SPU) with a clock speed of 725 MHz.
Each of the 1440 SPU can do 1 mad (multiply and add) operation per cycle (2 floating point operations per cycle)
So you have 1440 * 2 ops = 2880 ops per cycle.
Then, you multiply by the clock speed to get the flops (floating point operations per second).
2880 * 725 MHz = 2088000 Mflops = 2.088 Teraflops.

How to calculate theoretically possible GFLOPS
n0thing Apr 22, 2010 1:59 AM (in response to noxnet)Originally posted by: noxnet
I read that 5 SPUs build one 5DShader Unit, resulting in 288 5D Shader Units on an HD5850.
Can anyone explain this to me? What does 5D ShaderUnit mean?
Comparing SPUs to CUDA Cores?
A threadprocessor can execute upto 5 independent instructions simultaneously in 1 clock IF the shader compiler is able to find these instructions. Otherwise your processor would not be utilized of its full potential.

How to calculate theoretically possible GFLOPS
noxnet Apr 22, 2010 5:23 AM (in response to n0thing)Thanks for your quick replys!
So an HD5850 has 288 Thread Processors with 1440 SPUs in total.
How are these Thread Processors further splited into compute units? According to OpenCL queries an HD5450 with 80 SPUs has 2 compute units.
I guess concerning double precision 5 SPUs are needed for doing 1 dp calculation.
