Why include FP64

[FordGT90Concept]

What you pasted from Anandtech sounds like speculation to me. If they did indeed strip binary64 from the core FPUs then, to maintain backwards compatibility, they would need to have a scheduler that diverts binary64 workloads to the cores that can do it.

 

[WhoDecidedThat]

What you pasted from Anandtech sounds like speculation to me. If they did indeed strip binary64 from the core FPUs then, to maintain backwards compatibility, they would need to have a scheduler that diverts binary64 workloads to the cores that can do it.

NVIDIA GTX 680 Review: Fermi Distilled

 

[TheBrainyOne]

not everyone uses a gf cards sound controler and that uses space, should they chop that out for one more shader

Yes they should. It's a GPU. It's meant for graphics, not audio.

 

[FordGT90Concept]

NVIDIA GTX 680 Review: Fermi Distilled

Putting binary32 and binary64 in one FPU doesn't make much sense (that's the CPU approach where there's a lot more logic involved) so in GK104, they split them into separate FPUs altogether. The scheduler in GK104 must be able to decide which core to send the work to so obviously binary32 workloads go to the binary32 cores and the binary64 workloads go to the binary64 cores.

To your original question: they keep FP64 because there's demand (albeit small) for them.

The significance of GK104 is that they could relatively easily ramp up binary64 performance in the design if there is demand for it (like in Tesla products). In the Fermi design, ramping up binary64 would also ramp up binary32 which is a huge disadvantage if binary32 is not desired.

 

[TheBrainyOne]

they keep FP64 because there's demand (albeit very small) for them.

FTFY.