21 replies to this topic

[Jackal Noble]

LordNothing, on 07 September 2019 - 01:16 PM, said:

nah jut an understanding of fp maths. i write code where i need to make the fpu do things it really doesn't know how to do or even want to do. you can get 15 places but what if you need 20. the x87 instruction set used to have support for an 80-bit float but idk if thats still supported on 64bit machines (it would require a doubling up of registering operations as you need to load the extra 16 bits into a register). quadruple precision was supported on a lot of 70s/80s minicomputers, like the vax mainframe i used in highschool cs class (it was slow and your time was shared with other users).

but i guess this kind of stuff was deemed unnecessary for consumer hardware (its really only useful for the science and engineering sector). a lot of gpus still only support single precision or have reduced capabilities when using doubles, workstation gpus have the extra precision though, which is why they cost more. dp isnt needed for rendering though and is more of a hinderance, but if you want to do accurate photogrammetry operations on your gpu you might be better off with the more expensive card. though im not sure if they support quad out of the box (i doubt it). you can always gain extra precision in software though, even if you have to break out the large fixed point, and that usually comes with some performance penalties.

and really if i was that smart id be able to do these kind of maths in a way that requires less precision from the hardware.

That's about the best description on a workstation GPU I've seen, aside from actually looking into it. I still have a question mark over my head when I see the price of those things.

[LordNothing]

Jackal Noble, on 07 September 2019 - 04:11 PM, said:

That's about the best description on a workstation GPU I've seen, aside from actually looking into it. I still have a question mark over my head when I see the price of those things.

should also point out that double does come with some performance bottlenecks as you have to move twice as much data around. not that those cards cant do single, they just give you the option to have double. though i hope that wall comes down soon as id love to have gpu physics for large environments (like newtonian space flight simulators). you don't really have problems until you start operating small numbers with large numbers, but if you are trying to determine the result of an aerobraking maneuver where tiny amounts of drag are subtracted from a very large velocity vector. you could create a situation where the components you want to subtract are too small for the big number to even notice and end up getting rounded to oblivion and find yourself leaving the solar system. or worse end up with a nan error that crashes the whole thing (this can be very hard to debug, if you do something that you know could break the fpu its good practice to handle fp error exceptions).