Single Precision Floating Point in GPUs?

[SinOfLiberty]

I wanted to know more about what it does or its use is in GPU.

Maxwell has 13
Kepler has 8

 

[Snowman_]

Edit to remove false info.

 

[NTMBK]

"Single precision" is defined by IEEE standard to be 32 bits; "double precision" is 64 bits. So Maxwell has 13... what? Single precision FP ALUs per shader core? I'm not entirely sure.

 

[FatherMurphy]

Maxwell has 13
Kepler has 8

Is it possible he is referring to the GFLOP/s? Per NVIDIA's whitepaper on GM107 (http://international.download.nvidi...tional/pdfs/GeForce-GTX-750-Ti-Whitepaper.pdf), the GK107 does 812.5 GLOP/s and GM107 does 1305.6. Those numbers partially correlate to the numbers quoted. Though OP appears to have started this thread with an incomplete post.

 

[SinOfLiberty]

A floating-point number is basically a fraction. Precision is how many digits after the decimal place are used. Think of 1.25 vs 1.2567, the second number has a higher precision.

A GPU performs math calculations and at times it needs to do fractional calculations versus calculations with strict integers (1,4,9,7).

Now I'm guessing the number 8 for Kepler and 13 for maxwell refers to how many bits (bit is a binary digit) are reserved to keep track of the fractional portion of the number. The more bits reserved for the fractional portion leads to a much higher precision since each extra bit essentially doubles the precision.

I hope I stated this all correctly, I just woke up and typed this on my phone. You can get much more technical but I feel this is a decent overview.

http://en.m.wikipedia.org/wiki/Single-precision_floating-point_format

Thanks, this explains it nicely!

I guess Maxwell should eat Kepler in this area by quiet a bit.

 

[naukkis]

A floating-point number is basically a fraction. Precision is how many digits after the decimal place are used. Think of 1.25 vs 1.2567, the second number has a higher precision.

Just wrong. Floating point is what it describes, a floating point.

On integer calculations you have to decide decimal point place. That's fine as number scales are about equal, like 2.5000*2.5000 = 6,2500, there's enough precision to handle all information.

With floating point you have multiplier and exponent, like 2,5*10^25 or 2,5*10^-25. You have ability to handle much smaller and much bigger values compared to that presented without exponent.

As those examples:

2500000000000000000000000000000000000000 and 0,00000000000000000000000000000000000025

see how decimal point is floating......

 

[BrightCandle]

Its got to be referring to some sort of performance measure because a single precision float is 32 bit, defined by standard. There is just no way they are using 8 bit floats, that wouldn't be useful for anything to do with graphics. They are just 32 bit or occasionally 16 bit as that is what used to be present in the earlier GPUs (although I doubt many games use 16 bit anymore).

Its got to be referring either to some count inside of a core or the overall throughput.

 

[Loki726]

As mentioned above, a single precision floating point value is defined by the IEEE 754 as a 32-bit value that can be manipulated with a specific set of operations, such as multiplication, addition, etc.

In GPU specifications, single precision floating point operations per second is commonly referred to with the acronym FLOPs (floating point operations per second), or commonly GFLOPs (giga floating point operations per second), and is a measurement of throughput or operations per second (1GFLOPs is equivalent to 1 billion FLOPs).

Many GPU applications are limited by this throughput, which is why it is often quoted in specifications. The idea is that more FLOPs can roughly indicate better performance.

It is important to not read too much into this though. FLOPs are typically specified as a theoretical limit, rather than a rate that most or even any applications should expect to achieve. It is possible for a GPU to have a higher theoretical limit, but actually achieve less on average.
Furthermore, there are many other factors (literally hundreds) that affect performance, including the characteristics of the application itself.

 

[Snowman_]

Just wrong. Floating point is what it describes, a floating point.

Thank you for politely pointing out me being wrong...I must admit I feel incredibly embarrassed. I literally had my first formal lecture on the topic this past Wednesday, guess I need to re-read my notes...we went over an ALU that handles FP operations and I guess I didn't grasp it all trying to just keep up with the hardware.

Again thank you for the kind correction, I'll delete my post as it is clearly incorrect. Learn more each day 🙂

Sorry OP, was just trying to help.