Radeon Pro Duo Announced: Dual Fiji Graphics Card

[adakite]

Are your serious here? Yes 64 bit floats are important but to say nobody uses 32 bit floats seriously is asinine.

I'm not saying nobody uses FP32. What I meant is that computing (i.e., not gaming or photoshop effects) demands FP64.And so, I strongly request to have both FP32 and FP64 capabilities when talking about GPGPU. That's it.

[adakite]

And this statement is based on what?
Tonga-based FirePros, the entire series of Maxwell-based Tesla and Quadro cards have all terrible FP64 performance and still are used for professional computing. Hawaii is currently the only chip providing a 1:2 rate of dual to single precision performance, others are 1:16 or even worse. Kepler had 1:3, Tahiti 1:4 after all, I guess.
If only FP64 matters, why are so much Nvidias used and not Hawaiis all over? Hawaii provides almost twice the FP64 performance of GK110B (their best chip for FP64), power consumption is comparable, so almost twice the TFLOPS/Watt.
Newer GPUs (Polaris, Vega, Pascal) should support 2:1 rates, though. Nvidia implemented mixed precision mode for FP16/FP32 in their Tegra, I'm sure they are able to scale that up, like AMD already did with Hawaii.
Then you are free to talk about FP64

It seems people use computing for any "mathematish" used of their GPU. I'm using this term for scientific computing and so in this areas, FP32 is useless, period.
Nvidia offers a more mature environment that AMD, at least so far. Besides they built a huge community and academic support that is why (for now) nvidia is dominating HPC area.
At some point, AMD offers a great engine, but Nvidia has a better car if you look at multiple aspects.

But things are moving pretty fast and AMD is clearly moving forward with open-source in mind, which is nice, really nice. Their HIP compiler is a real killer.