When calculating the amount of FLOPS in a GCN GPU.

[nigelhere9901]

The 7970 has 32 GCN Compute Units, Each GCN Compute Unit contains 4 SIMDs 16-wide.

 

4 SIMDs 16-wide per each Compute Unit = 4 X 16 = 64 SIMDs per each GCN Compute Unit

 

64 SIMDs X 32 Compute Units = 2048 Stream Processors, correct?

 

2048 (stream processors) X 0.925 (frequency in GHz) X 2 (Operations per cycle, 2 FMA) = 3788.8 GFLOPS

 

 

 

What does the number 2 represent? Did I label all of them correctly?

Edited April 29, 2013 by nigelhere9901

[sivispacem]

It signifies two floating point operations per clock cycle. This integer varies depending on the processor design and type. I think the Intel Sandy and Ivy Bridge processors were 6 (most) or 8 (Haswell) operations per clock cycle per core, for example.

[nigelhere9901]

 

It signifies two floating point operations per clock cycle. This integer varies depending on the processor design and type. I think the Intel Sandy and Ivy Bridge processors were 6 (most) or 8 (Haswell) operations per clock cycle per core, for example.

Thanks for the reply, been wondering about that.

 

Also, do you happen to know how to calculate FLOPS in a CPU? Is it the same method? Cores * Speed in GHz * FLOPS per cycle?

Edited April 29, 2013 by nigelhere9901

[sivispacem]

Technically, it's Sockets * Cores Per Socket * Clock Speed Per Core * Floating-Point Operations per-cycle, but the former number is usually 1 in most desktop systems.

 

The actual calculation for the 7970 would be 32 * 64 * 0.925 * 2 = 3788.8 GFLOPS

[nigelhere9901]

The actual calculation for the 7970 would be 32 * 64 * 0.925 * 2 = 3788.8 GFLOPS

That theory is not any different from mine, each Compute Unit has 64 stream cores, you just have to multiply the number 64 to the number of compute units and then in to speed and then into Floating Point Operations per cycle.

[sivispacem]

The mathematics is the same, but it's showing the process in the main sum. That way you don't have to work out total core numbers, which is relatively okay on a single GPU or server but much more difficult across distrusted computing systems.

[nigelhere9901]

Alright, Thanks for your help kind sir.