Double-Precision and Floating-Point Precision?!

[ClobberXD]

These are very common terms in the world of Graphics Processing: Double-Precision, and Floating-Point Precision. Can someone explain to me what they are, and how are GPUs with these, used differently from the ones without?

 

Thanks!

[ARikozuM]

https://en.wikipedia.org/wiki/Double-precision_floating-point_format

The article says it pretty well.

[Nineshadow]

float and double are floating point types used to store real numbers. You can think of them as a scientific notation.

They are called floating point because the decimal/binary point can be placed anywhere relative to the digits of the number, a position which is indicated by the exponent.

Of course, since we're talking about computers, these numbers are represented in binary. And here we have a problem, because there isn't a way for a computer to be 100% accurate without having infinite memory to store the numbers. Just like how in base 10, 1/3 = 0.3333333333... , there are similar recurring fractions in binary.

 

That's where the memory allocated to one floating point number comes in. The more memory it has, the more digits it can represent, and more accurately can it depict real numbers.

 

Most commonly, floating point numbers are 32 bit. That's the standard nowadays. For use cases when you need more precision, we use 64-bit floating numbers, which are called double-precision floating point numbers(for pretty obvious reasons 2*32=64). Of course, this also allows for larger numbers as well.

 

[ClobberXD]

6 minutes ago, ARikozuM said:

https://en.wikipedia.org/wiki/Double-precision_floating-point_format

The article says it pretty well.

But I still don't understand how it's very useful, that Autodesk recommends getting only Quadros and FirePros... Can you tell me how it affects performance?

[ClobberXD]

3 minutes ago, Nineshadow said:

float and double are floating point types used to store real numbers. You can think of them as a scientific notation.

They are called floating point because the decimal/binary point can be placed anywhere relative to the digits of the number, a position which is indicated by the exponent.

Of course, since we're talking about computers, these numbers are represented in binary. And here we have a problem, because there isn't a way for a computer to be 100% accurate without having infinite memory to store the numbers. Just like how in base 10, 1/3 = 0.3333333333... , there are similar recurring fractions in binary.

 

That's where the memory allocated to one floating point number comes in. The more memory it has, the more digits it can represent, and more accurately can it depict real numbers.

 

Most commonly, floating point numbers are 32 bit. That's the standard nowadays. For use cases when you need more precision, we use 64-bit floating numbers, which are called double-precision floating point numbers(for pretty obvious reasons 2*32=64).

 

They can be more precise alright, but how is it useful, in say, 3D-Rendering applications?

[ARikozuM]

I'm not sure exactly [on how to explain it].

In engineering, you're only allowed so much margin for error. As a result, the double precision is necessary to keep the tolerances very tight. Imagine drawing two lines on paint. You zoom in and they start to separate. If you do the same on Autodesk, the line will still be connected and if you placed an angle indication, it will continue to be as precise as possible.

 

Come to think of it, two applications where it will matter a lot is prosthetics where having exactly the right amount of material is important for their use and micro-machinery where using scales doesn't work well with computers.

[Imakuni]

5 minutes ago, Anand_Geforce said:

But I still don't understand how it's very useful, that Autodesk recommends getting only Quadros and FirePros... Can you tell me how it affects performance?

Because those GPUs have drivers specifically tuned for such apps. They also go through more extensive validation than regular consumer cards to assure stability, as well as next level customer support. That's why they are the only ones recommended.