FP16 tensorcore = 312 tflops x 2 (counting sparse as dense) = 624 tflops x 8 GPUs = 5 "pflops" The missing 8x you are looking for is just because tensorcore

FP16 tensorcore = 312 tflops x 2 (counting sparse as dense) = 624 tflops x 8 GPUs = 5 "pflops" The missing 8x you are looking for is just because tensorcore math is much faster than their normal fma path.