Hi David,

Nice article, i like to do similar comparisons myself from time to time.

I would like to add that for MOST commercial applications SP computations is a viable option. DP is most often a scientific research requirement ( some people will disagree here but that's because they dont know how analyze their code... :-).

Hence an SP/mm2 and SP/watt comparison is also very relevant.


For example the GTX680 does about ~16 GFLOPS/watt SP, which is excellent for both signal and image processing.

The 640m (GK107 version) does 492 GFLOPS at 20 watts => 24.6 GFLOPS/watt! (extremely good!)

On another note i do think you might want to include the latest AMD GCN architecture in the comparison as well.

About utilization I would say that GPUs generally reach higher utilization with less effort than in CPU programming. What people forget to do is to look at the correct metrics, achieved bandwidth utilization or raw FLOPS utilization, where people have a tendency to just look at the FLOPS. Actually many applications become bandwidth bound on the GPU due to the much higher GFLOPS / bandwidth ratio.

So if you have a bandwidth bounded application ( ie mainly moving data) you need to measure utilization in achieved application bandwidth / theoretical bandwidth. This is what most people dont figure out.

The answer becomes clearer when you begin looking at the GFLOPS/bandwidth ratios.

The idea that it's easier to squeeze more theoretical performance with multi-threading and SSE instructions on a CPU is unfortunately not true.

Thank you,