According to data provided by Indiana University, the Big Red 200 supernumerary initial planning computing performance is 5.9P Flops (590 million floating-point computations per second), using the Tesla V100 acceleration card, but upgrading the next NVIDIA GPU to 8P Flops, and the number of GPUs required is even smaller, because the new architecture (that is,7nm Ampere) can bring 70-75% performance improvement!
70% of the performance improvement is already strong, but this may not yet be the real strength of 7 nm amperes, and the results calculated by twitter user dylan 522p are even more surprising because the 7 nm ampere gpu has improved its performance by more than 100%.
According to his results, FP64 performance of 7 nm Ampere GPU up to 18 TFLOPS As a contrast, the Tesla V100 Acceleration Card is 7-7.8 TFLOPS and the latest Tesla V100s is 8.2 TFLOPS.
Among the Tesla series accelerators, fp32: fp64 is 2:1, that is to say, the fp32 performance of 7Nm amp graphics card can reach 36tflops, which is more than twice the current performance of my Tesla V100 series, and the latter is the strongest but 16tflops.
From these information, the performance improvement of NVIDIA's next generation GPU - 7Nm amperes will be very large. The increase of more than 100% is far greater than the performance improvement of Pascal to Volta's generation by 50%. The latter is only from 21tflops to 30tflops. After the 2000's, the performance improvement of GPU by multiple times is very rare.
Of course, how game-level 7nm Ampere ended up is hard to say, because NVIDIA's recent generations of GPU Tesla and Ge Force series are separate, RTX graphics card for the 12-nm Turing GPU full version of only 4608 CUDA core, Volta is 5120, and game GPU's FP64 performance castrated to 1, comparable only FP32:32 performance, TU core with GV100 core FP32 performance almost all around 14-16 TFLOPS.