Valued Gamer
Banned
I have been wondering what the efficiency gains have been from RDNA 1 to RDNA 2, not just in a power per watt sense, but comparing an RDNA 1 tflop to an RDNA 2 tflop.
AMD have said that RDNA 1 gave a 50% improvement in performance per watt over GCN.
Digital Foundry did some benchmarking of GCN cards vs RDNA 1 cards and found a 27% performance improvement in a RDNA 1 flop over a Polaris GCN flop.
Now looking to this generation, AMD have again said that RDNA 2 gives over 50% performance per watt improvement over RDNA. That's very impressive considering the jump in performance per watt from GCN to RDNA 1 included a node shrink, whereas the jump from RDNA 1 to RDNA 2 doesn't have that same node change.
So are we able to work out what the performance improvement per flop will be with RDNA 2 over 1?
Can we say that because there was a similar performance per watt again, that we can expect a similar performance per flop as well?
If so that gives us a 50% improvement over GCN
So if we look at the PS5, it would be like a 15tflop GCN GPU.
So are we right to assume that?
Is there any other way we can work out a performance per flop increase using the information AMD has released?
AMD have said that RDNA 1 gave a 50% improvement in performance per watt over GCN.
Digital Foundry did some benchmarking of GCN cards vs RDNA 1 cards and found a 27% performance improvement in a RDNA 1 flop over a Polaris GCN flop.
Now looking to this generation, AMD have again said that RDNA 2 gives over 50% performance per watt improvement over RDNA. That's very impressive considering the jump in performance per watt from GCN to RDNA 1 included a node shrink, whereas the jump from RDNA 1 to RDNA 2 doesn't have that same node change.
So are we able to work out what the performance improvement per flop will be with RDNA 2 over 1?
Can we say that because there was a similar performance per watt again, that we can expect a similar performance per flop as well?
If so that gives us a 50% improvement over GCN
So if we look at the PS5, it would be like a 15tflop GCN GPU.
So are we right to assume that?
Is there any other way we can work out a performance per flop increase using the information AMD has released?