OMG FidelityFX Super Resolution 2.0 (FSR 2.0) The technology is arriving later this year, but before it drops, AMD has outlined exactly what makes its super upscaling algorithm work. A presentation at GDC outlined what has improved with FSR 2.0 as well as what has been optimized, and this is good news for RDNA 2 players in this regard.
AMD’s latest GPU architecture features some exciting new optimizations to run FSR 2.0.
There are a lot of improvements coming with FSR 2.0 – an hour-long GDC conversation could just cover it at a higher level. But generally, you’re looking at improved image quality, speed, antialiasing, and tons of optimizations.
These improvements come in all shapes and sizes: tonemap optimizations to free up GPU resources, Cache Blocking to reduce time spent looking out of local memory, and reworks of the FidelityFX Single Pass Downsampler to aid in the low overhead auto-exposure feature, to name a few. There’s also a new feature to “lock” wide pixel ridges in place to ensure thin features don’t flicker or disappear on screen, like fences or distant objects.
All of this will work on AMD, Nvidia, and even Intel’s upcoming GPUs. FSR and FidelityFX features are generally open source.
Although there are also specific optimizations based on RDNA-2 to improve FSR 2.0.
For example, there are specific RDNA-2-based enhancements to Lanczos approximation, an underlying algorithm for FSR. RDNA 2 shaders can also use Wave32 or Wave64 modes. While FSR shaders run in Wave32 mode most of the time, they can sometimes benefit from Wave64, which means RDNA 2 GPUs can capitalize on it at times.
RDNA 2’s Infinite Cache also seems useful in some cases here – FSR 2.0 can get very cache hungry with 4K frames. Even with a larger cache, AMD says some leaks can occur. So it’s using Cache Blocking to help alleviate congestion by splitting the workloads and running them all in the local area cache to increase hit rates. This should also benefit other GPUs without as much local memory.
And just because AMD is rolling out improvements to RDNA 2-based GPUs like the RX 6800 XT, it’s ensuring that those improvements don’t get in the way of older Radeon GPUs or the competition, he says.
To do this, there are several fallbacks in place. Optimizations that might get in the way of non-RDNA-2 cards are flagged and FSR 2.0, when properly configured by AMD, should automatically ignore them when running on other GPUs. Wave Size is the example AMD gives, which is only configured to use Wave64 when FSR performance actually increases on suitable GPUs as a result.
FSR 2.0 looks like a huge leap forward in AMD’s upscaling technology, and I hope it’s similar to Nvidia’s leap in performance and quality between DLSS 1.0 and DLSS 2.0. With that comes “significant” performance gains over FSR 1.0, although it also has its downsides. AMD says that FSR 2.0, unlike FSR 1.0, it is only recommended for use on very recent and powerful GPUs. The Nvidia GeForce GTX 1060, for example, is not recommended by AMD for use with FSR 2.0.
While some gamers with older technology may not benefit as much from FSR 2.0 as others, AMD is bringing FSR 2.0 for Xbox with Microsoft so console gamers can experience high-quality upscaling.