Friday, March 24th 2023
AMD FSR 3 FidelityFX Super Resolution Technology Unveiled at GDC 2023
AMD issued briefing material earlier this month, teasing an upcoming reveal of its next generation FidelityFX at GDC 2023. True to form, today the hardware specialist has announced that FidelityFX Super Resolution 3.0 is incoming. The company is playing catch up with rival NVIDIA, who have already issued version 3.0 of its DLSS graphics enhancer/upscaler for a small number of games. AMD says that FSR 3.0 is in an early stage of development, but it is hoped that its work on temporal upscaling will result in a number of improvements over the previous generation.
The engineering team is aiming for a 2x frame performance improvement over the existing FSR 2.0 technique, which it claims is already capable of: "computing more pixels than we have samples in the current frame." This will be achieved by generating a greater number of pixels in a current frame, via the addition of interpolated frames. It is highly likely that the team will reach a point in development where one sample, at least, will be created for every interpolated pixel. The team wants to prevent feedback loops from occurring - an interpolated frame will only be shown once, and any interpolation artifact would only remain for one frame.
However, a number of potential setbacks were noted - a reliance on color clamping to correct color of outdated samples is not entirely feasible. It will be difficult to produce non linear motion interpolation on 2D screen space motion vectors, and the interpolation of final frames will mean that all post-processing needs to be interpolated, also counting the user interface in the foreground. One of AMD's diagrams shows how a native rendering technique stacks up to FSR 2.0 and 3.0.
FSR 3.0 will enable a smoother overall gaming experience, and simultaneously this allows developers to focus more GPU time on visual quality. Latency reduction is a key focus area for FSR 3.0 - AMD has the gamer in mind, with high frame rates and the lowest achievable latency as basic requirements. The engineers are also aiming for a smooth upgrade path from titles that currently utilize version 2.0 of FSR.
Source:
GPU Open Press Material
The engineering team is aiming for a 2x frame performance improvement over the existing FSR 2.0 technique, which it claims is already capable of: "computing more pixels than we have samples in the current frame." This will be achieved by generating a greater number of pixels in a current frame, via the addition of interpolated frames. It is highly likely that the team will reach a point in development where one sample, at least, will be created for every interpolated pixel. The team wants to prevent feedback loops from occurring - an interpolated frame will only be shown once, and any interpolation artifact would only remain for one frame.
33 Comments on AMD FSR 3 FidelityFX Super Resolution Technology Unveiled at GDC 2023
fuckkkkk, i hate this trend
also so Fluid motion is back so video playback will change to 60fps?
If not, guess I will be using Vega cards/older cards forever and proabldy end up selling my 5000/6000 GPU's.
the previous gen GPU's might only get an improvement but that is only on what FSR 2 can offer. Just don't get your hopes up I am just giving you a realistic idea on what you need to expect on FSR 3.
If you can recall FSR 3 was introduced in RDNA 3 launching, so those new feature that they were talking about performance uplift were run on RDNA 3 GPU.
So that explains why RDNA3 GPU are expensive because they added 2 AI Accelerators 1 RT per CU to prepare this Hardware on FSR 3.You shouldn't, don't join with this Ray Tracing bandwagon thing... just run your game without ray tracing and you will be good for a couple of years. I watched my son play games on his new computer with Radeon 6000 card, disabled ray tracing and still appreciate how crisps the graphics are. So yeah you will be fine without ray Tracing...
Now AMD is following suit with all of Nvidia's garbage!
How about create cheaper and faster GPU's? Jesus Christ, how did we get from 25-50% faster performance for the same price generation over generation to 0-50% faster performance at 50% higher price!
DLSS frame gen introduces latency and artifacts. HWUB states that the tech is only really useful between 70 - 120 FPS and even then, only in certain titles.
You are getting more frames but you can't interact with any of those frames. It's nice to look at but horrible disorienting if your latency is too high as the disconnect between what you see and you inputs grows.
People need to actually try it not just watch videos of it which to me are useless especially with how youtube handles above 60fps.
I don't even actually need it for good performance in the two games I like it in but still choose to use it.
It is useless for MP game which I do play alot though. But I will probably try it in D4 when it comes out just to see the implementation.
So meh on the announcement. I'll wait until a real review is done.
FSR 3 will land, and despite IQ shortcomings, it will be reviewed as a DLSS3/FG killer, and virtually overnight a whole demographic of PC gamers will suddenly think it's a killer tech and since it's acceptable IQ wise, and open so DLSS3/FG should go die in a fire.
Well we already know it's a great and desirable tech, and the more competition the better, that's good for everyone.
Another massively important footnote, it's optional, so if the mere thought of 'fake frames' is a line you've decided should not be crossed, fear not, just don't enable it. Bonus point! all frames are fake frames.
Honestly, I'm somewhat excited for this one, the architecture looks like a meaningfully different take compared to DLSS 3, unlike FSR 2 which whilst I applaud the competition, neither did I really care as an Nvidia user already being serviced by DLSS 2.
Maybe in GPU architecture we might be waiting to 2026 or more to AI added inside it. AI will be part of CPU architecture for first.
If you render 60 frames per second, the difference between frame 13 and 14 is very small, AI with motionvectors could easily make an educated guess of what the frame in between would look like, so really it is a fine tech.
I just dont want, same for the upscaling techs, that the devs of games become lazy when it comes to optimizing or that the gpu's advance less because this tech will solve all the performance lacking problems (while obviously costing the same if not more.....)
Nvidia has their tensor cores which are AI accelerators in their geforce gpus. Amd has some AI hardware in RDNA3 gpus now too.
From an architectural standpoint, today's ML cores have much more in common with GPU raster cores than CPU cores.
If you're waiting for AI to be included in CPU architecture, be patient because you may have to wait a long time. A really looooooong time.
Suit yourself. The 3070ti at 100 fps had lower latency (by 50%!!) than the 6700xt at 200 fps. That's absolutely insane