The Student Working On "Soft" FP64 Support Is Good News For Older GPUs

[fuzz]

I am sure this library can also be useful for nouveau too, to help provide GL4 support for those NV cards that don't have hardware FP64 support. This will help certain older cards work with GL4 when it comes time.

The most ironic thing to me is older NVIDIA cards would be able to run better under open source drivers than closed source drivers. Joke's on you, NVIDIA sabotage.

[Mez']

Good news. While forcing OGL version works most of the time, if you run into an unrelated bug and want to take a closer look, a forced context can be annoying, as things like apitrace don't work properly with them. Having that library just for this reason is a good enough reason in my opinion. Speed doesn't matter, there are still 0 games using fp64.

Most of the time, yes.
But still I ran into an annoying bug in Ubuntu 3-4 months ago which made the override option not to work anymore (and side effects). Neither Dirt Showdown nor any other game requiring the override to 4.1 would work. A Steam or xserver-xorg-video-ati update eventually solved it 2 months later.

If this work is integrated to mesa, then this kind of bug can be avoided without having to manually override the GL and GLSL versions. This has been awaited for some time, hence I'm just glad it's on its way.

If everybody gets to thank him in french:
Bon boulot! Ça faisait un moment que les détenteurs de cartes un peu plus anciennes (mais toujours en bon état) attendaient ça! Merci.

[gsedej]

Hi,

can someone explain me about FP64 performance. My collages are researchers and they develop algorithm that need doubles. Single floats are not good enough (the data would corrupt too much)...

Anyway we NEED doubles

There are lists of AMD and NVIDIA gpus on Wikipedia. The gpus have Single percision and Double percision score in GFLOPS. All nvidias with Maxwell and Pascal has BAD double performance - 1/32 (3%) of single (except 5000$ "P100"). Even freaking Nvidia Titan X has 317GFLOPS in FP64 (vs RX 480: 323)
The AMD has bit better ratio, mostly 1/16 (6%), or even better in the top class.

Does this mean that 99% of GPUs have "soft FP64"? Or simply very small HW part for double arithmetics?
Also why doesn't lower GPUS (R5, HD x6xx or less, ...) mentioned GFLOPS for doubles?

[Tomin]

Does this mean that 99% of GPUs have "soft FP64"? Or simply very small HW part for double arithmetics?

Only small portion of the GPU is capable of doing double precision calculations. Nvidia thinks that those who need double precision performance should take a look at Tesla cards.

Also why doesn't lower GPUS (R5, HD x6xx or less, ...) mentioned GFLOPS for doubles?

They do not have double precision hardware at all, I think. It's not relevant for computer graphics usually and there are no games that need it.

[smitty3268]

Hi,

can someone explain me about FP64 performance. My collages are researchers and they develop algorithm that need doubles. Single floats are not good enough (the data would corrupt too much)...

Anyway we NEED doubles

There are lists of AMD and NVIDIA gpus on Wikipedia. The gpus have Single percision and Double percision score in GFLOPS. All nvidias with Maxwell and Pascal has BAD double performance - 1/32 (3%) of single (except 5000$ "P100"). Even freaking Nvidia Titan X has 317GFLOPS in FP64 (vs RX 480: 323)
The AMD has bit better ratio, mostly 1/16 (6%), or even better in the top class.

Does this mean that 99% of GPUs have "soft FP64"? Or simply very small HW part for double arithmetics?
Also why doesn't lower GPUS (R5, HD x6xx or less, ...) mentioned GFLOPS for doubles?

Traditionally FP64 is a feature used by compute workloads, and not by games. As you mention with fp32 the data can be too inexact for scientific workloads, while in a game the human eye is incapable of noticing a color that's been rounded incorrectly.

So, NVidia and AMD have worked to segment their hardware sales, to make lots more money by selling "professional" cards like Tesla that provide good fp64 support. Then they provide the slow 1/32 performance in their normal line of cards primarily so that anyone can test and develop fp64 there, even if they don't have great performance.

In a few cases, this limitation was done artificially, and the underlying hardware is actually there and just limited by software (BIOS or drivers). In other cases, the hardware is actually missing the extra fp64 hardware (an optimization which lets them pack in extra fp32 hardware for the same cost).

The older cards which don't list a speed are the ones without hardware support. I'd expect the soft float performance would be quite slow though. Presumably it could vary depending on the exact workload and mix of operations you were using.