Linux "RADV" Radeon Driver Gets A Big Speed-Up For 16-bit FidelityFX Super Resolution

[dimko]

What is this "16-bit FSR" referring to?
If I have screen/GPU that work on 24bit colour, does benefit of 35% in any way touch me?

[qarium]

What is this "16-bit FSR" referring to?
If I have screen/GPU that work on 24bit colour, does benefit of 35% in any way touch me?

if you see the AI/DLSS hardware in nvidia hardware it is 4bit/8bit/16bit/32bit mix. (nvidia hardware can also do 64bi5 of course)
amd RX580 only has 32bit and 64bit and my vega64 has 16bit and 32bit and 64bit

it does not mean that the hardware represent 24bit color in only 16bit with less quality.

No it means some part of the AI agloritmus need to be calculated but it does not need to be calculated at 32/64bit. some parameters in DLSS only need 4bit precision. of 4bit is enough why do it in 64bit ?

and these vector/simd units are made in a way that you can do the double amounts of single calculations if you do 16bit instead of 32bit... if you only use 8bit you can do the double amount as 16bit
and if dome calculation only need 4bit precision you can do the double amount than 8bit.

just go back in history of the first intel cpu it was only 4bit... many calculations do not need more than 8bit.

this "benefit of 35%" by using 16bit floading point really means it is used in a part of the algorytm who does not need higher precision and the result is the same quality with 35% higher performance.

on the same hardware DLSS2.3 is 6% faster at the same result than FSR2.0 at 16bit FP because it use 4bit and 8bit calculation in moments no higher precision is needed.

You have to unterstand the performance does not come out of a black-hole vaccum instead it comes from more calculations per second and on the same transistor amount you can only do this with lower precision like 4bit or 8 bit or 16bit. and in many moments in these AI and algoritms there is no need for higher precision.

just imagine this: theoretical you would even get more performance if you add 2bit compute to with only 2 bit you can only represent 4 numbers but in many cases you do not need higher precision...

2BIT=4 numbers
4bit=16 numbers
8bit=246 numbers
and so one and so one. many calculations do not need more than this..

[dimko]

if you see the AI/DLSS hardware in nvidia hardware it is 4bit/8bit/16bit/32bit mix. (nvidia hardware can also do 64bi5 of course)
amd RX580 only has 32bit and 64bit and my vega64 has 16bit and 32bit and 64bit

it does not mean that the hardware represent 24bit color in only 16bit with less quality.

No it means some part of the AI agloritmus need to be calculated but it does not need to be calculated at 32/64bit. some parameters in DLSS only need 4bit precision. of 4bit is enough why do it in 64bit ?

and these vector/simd units are made in a way that you can do the double amounts of single calculations if you do 16bit instead of 32bit... if you only use 8bit you can do the double amount as 16bit
and if dome calculation only need 4bit precision you can do the double amount than 8bit.

just go back in history of the first intel cpu it was only 4bit... many calculations do not need more than 8bit.

this "benefit of 35%" by using 16bit floading point really means it is used in a part of the algorytm who does not need higher precision and the result is the same quality with 35% higher performance.

on the same hardware DLSS2.3 is 6% faster at the same result than FSR2.0 at 16bit FP because it use 4bit and 8bit calculation in moments no higher precision is needed.

You have to unterstand the performance does not come out of a black-hole vaccum instead it comes from more calculations per second and on the same transistor amount you can only do this with lower precision like 4bit or 8 bit or 16bit. and in many moments in these AI and algoritms there is no need for higher precision.

just imagine this: theoretical you would even get more performance if you add 2bit compute to with only 2 bit you can only represent 4 numbers but in many cases you do not need higher precision...

2BIT=4 numbers
4bit=16 numbers
8bit=246 numbers
and so one and so one. many calculations do not need more than this..

This is why I love open source communities...