For the last time: "FLOPSsp ≈ f × n × 2 (FMA)"=wrong, does not exist. Correct="FLOPSsp ≈ f × n × 2 (FMA2)". Fermi="FLOPSsp ≈ f × n × 3 (FMA3)". They don't write it because the third OP is not certain, but with a good driver is above 90% certain for games. They actually explain that in the past, and in g80-Gtx200 models they actually count it "FLOPSsp ≈ f × n × 3 (madd+mul). Second: gtx580=1.6TflopsSP is in 64bit Dual-Isue. Fermi cores are 64bit, you can't compare them with 32bit cores of Radeon6970=2.7Tflops. You must convert Fermi Flops in 32bit if you want to be accurate, thats 2x tested. Anyway Fermi is faster in 99% of all benchmarks, I don't thing you have an objection here, are you? Ad at the end, a card with 3 billion email@example.com, can't lose to a card (same generation) with 2 billion transistor@880mhz. And Kepler with 3 billion transistor@3ghz, 640cores-128bit-quad-issue(probably not sure), will be 5 times faster than gtx580. Also don't forget all the Radeon problems on Linux.