Why don't graphics card's performance irl coincide with the FP32 compute performance they are rated at?
For example
R9 Fury X: 8.2 TFLOPS
GTX 980: 5.6 TFLOPS
Rx 480: 5.2 TFLOPS
GTX 1060: 3.9 TFLOPS
A lot of times the GTX cards beat their AMD counterparts despite having lower FP32 compute performance. Why is that?
I asked this because FP32 compute performance has always been my guide to buying graphics card but now I'm not so sure anymore.
>>56035236
FP32 is one type of processing that the graphics card do.
The reason that nvidia cards can beat them is because they have tons of processing extensions that are based around gaming. They also tend to have a lot better drivers (overhead wise, etc).
>>56035283
>FP32 is one type of processing that the graphics card do.
Isn't this the main kind of processing power used though?
>The reason that nvidia cards can beat them is because they have tons of processing extensions that are based around gaming.
Then why do they perform so badly on modern APIs like DX12? I know there's only like 10 DX12 games but still.
>>56035236
AMD has shit drivers. Because Pajeets can't write any good ones. And AMD only hires Pajeets.
That's it. That's really it.
pixel fillrate, rasterization, etcetera optimization.
It doesn't matter how many TFLOPS you have if you can't render stuff fast enough.
As it stands right now AMD is one gen behind Nvidia when it comes to optimization and is very likely AMD will never catch up.
AMD decided to re brand the 200 series instead of improving upon it unlike Nvidia that came out with Maxwell once they realized we were going to be stuck on 28nm for another 2 years as 20nm was a complete failure.
Any workstation user that doesn't need CUDA (aka Karlie Kloss Kode Kit) buys AMD for the shear high precision shader horsepower.
Nvidia = Gamer Babbys
AMD= Workstation Gurus
>>56035413
Because those older cards werent designed to take advantage of it. Look at Pascal, it has huge DX12 improvements versus cards that are equal to it in dx9/11