CEO of AMD saying RDNA TF > GCN TF in Navi conference.
Some random tweet: TF is TF.
No she didn't say that.
She said RDNA realises better gaming performance per clock cycle than GCN. The marketing material calls it IPC even though it isn't strictly speaking measuring instructions per cycle.
I've explained this before.
A FLOP is defined as a Floating Point Operation.
A Float or floating point number is a number with decimal places (e.g. 4.35) as opposed to an integer (e.g. a whole number such 4).
A floating point operation is quite literally performing a mathematical operation (i.e plus, minus, divide, multiply or indices) on a floating point number.
A TeraFLOP is 1 trillion floating point operations per second. One architectures floating point operation cannot by definition be more or less than another floating point operation.
The only thing that can change is precision (FP16/32/64) or how many decimal places the GPU calculates to.
The thing is, gaming performance is not directly measured by how many floating point operations a GPU can perform. There is a general trend in that more flops means more performance, but it is not a direct relationship. There are other factors at play. Stuff like geometry performance, pixel fill rates, raster operations and so on. Even cache hierarchy and bandwidth impact performance in terms of how quickly the GPU core can be supplied with numbers to crunch. RDNA offers specific improvements in terms of gaming performance but also offers some advantages in pure compute due to improved caching and of course, higher frequencies.
But if you stick an RX5700 up against a Radeon Pro WX9200 (i.e Vega 64) in a pure computational workload, then the latter will win, because it can execute more floating point operations than the former.
Why? Because a FLOP is a fucking FLOP.
Jesus fucking Christ.