How does FSR3 work on so many GPUs?

[Gat Pelsinger]

It's crazy how FSR3 with its improved upscaling and frame generation, a.k.a Fluid Motion Frames, support so many GPUs from various generations, and even the GPUs of it's one and only competitor, Nvidia! But how did AMD achieve this? AI (and also ray tracing) needs dedicated hardware acceleration to achieve acceptable performance. If older GPUs don't even have any hardware acceleration, how does AMD not run this AI compute tasks on the generalized graphics compute cores and actually degrade performance? Same happened or still happens with Intel's XeSS upscaling on older GPUs. They go from playable fps to actually unplayable fps, because they probably can't use the hardware acceleration built into different graphics cards.

 

A guess I would make is that these AI tasks take a very moderate number of cores and processing power, and the output is way more than what you would get by saving that processing power for traditional rendering. But then this contradicts my Intel example above, so either maybe AMD made it way more efficient than Intel, or Intel was suffering from a different problem.

[emosun]

they programed it to. thats how.

much in the same way a console from 18 years ago can play gta5

[filpo]

9 minutes ago, Hensen Juang said:

It's crazy how FSR3 with its improved upscaling and frame generation, a.k.a Fluid Motion Frames, support so many GPUs from various generations, and even the GPUs of it's one and only competitor, Nvidia! But how did AMD achieve this? AI (and also ray tracing) needs dedicated hardware acceleration to achieve acceptable performance. If older GPUs don't even have any hardware acceleration, how does AMD not run this AI compute tasks on the generalized graphics compute cores and actually degrade performance? Same happened or still happens with Intel's XeSS upscaling on older GPUs. They go from playable fps to actually unplayable fps, because they probably can't use the hardware acceleration built into different graphics cards.

 

A guess I would make is that these AI tasks take a very moderate number of cores and processing power, and the output is way more than what you would get by saving that processing power for traditional rendering. But then this contradicts my Intel example above, so either maybe AMD made it way more efficient than Intel, or Intel was suffering from a different problem.

Because for nvidia's gpus it depends on the hardware and AI accelerators in the actual die/on the board of the gpu but with fsr 3 it is based on an engine level 

 

But it does still need some AI accelerators to make it work so FSR 3 frame gen only works on 20 series or 5700 series and up (which you might not have picked out)

FSR 3 upscaling does work on 10 series nvidia cards and amd rx 590s and above tho

[jaslion]

1 hour ago, filpo said:

But it does still need some AI accelerators to make

The 5700 doesn't have any ai accelerators. The 6000 series does neither however they have dedicated ray accelerators.

 

So I assume it's basically a case of the architecture used happens to be compatible with full fsr 3 on the 5700 and the rx 6000 series has the architecture + dedicated cores so that might push further eventually.

 

This does also bring up the idea of potentially backporting full fsr 3 to older cards as it seems to be able to run on non ai and non rt core dedicated cards.

[Sharkyx1]

On 9/30/2023 at 3:30 AM, jaslion said:

The 5700 doesn't have any ai accelerators. The 6000 series does neither however they have dedicated ray accelerators.

 

So I assume it's basically a case of the architecture used happens to be compatible with full fsr 3 on the 5700 and the rx 6000 series has the architecture + dedicated cores so that might push further eventually.

 

This does also bring up the idea of potentially backporting full fsr 3 to older cards as it seems to be able to run on non ai and non rt core dedicated cards.

Amd wrote the  software as an async compute task, it doesn't have any ai processes involved. It's in short, a system of combining frames, nvidia includes hardware to accelerate the process, AMD on the other hand wrote code, that hypothetically can use some of the rendering time on some of the hardware that's not being utilized to complete the process. There are advantages to both methods.