can anybody help me whether there is a page of benchmarks of different GPUs available where I can see how fast a model is executed on different GPU-s of nvidia? (E.g.: I would like to know how much the forward prop. path would take at a 3060 GPU vs 3090 GPU vs the new 4080/4090 GPU) - assuming that 3060 is 100%…or which formula would you recommend to calculate this?
I personally don’t have those kinds of benchmarks of GPU performance on consumer GPUs, but the benchmarking exists for prosumer/enterprise GPU. External references may be your best bet, like GPU Benchmarks for Deep Learning | Lambda.
Fun fact: it is difficult to do a standard correlation across classes of GPUs as there are many internal differences between the GPUs as they are purpose built with different amounts of processor types, so you can probably only reasonably estimate generationally. Example: I’ve seen a GV100 do in 3 seconds what takes an RTX8000 took 27 because the particular algo required heavy FP64 calculations.