@Anatole I think I might have found it:
On the contrary, I'm pretty sure he said it was very recent info. So maybe not to the day current but it at least seemed like he meant within the past month or so.
www.resetera.com
The post that I am imagining in my head estimated the time for DLSS to run on 2-3 potential Dane configurations in ms. TBH, I may be remembering a composite of various related posts.
With that said, I replied to this Thraktor post at the time, and the results still puzzle me. In theory, the cost of DLSS should marginally increase with input resolution, since the earliest layers in the neural network are at input resolution. However, that cost should be fairly negligible (less than 10% of the total cost in the Kaplanyan/Facebook paper) compared to the reconstruction cost at output resolution.
With that in mind, it doesn’t make sense to me that the DLSS compute time should decrease as input resolution increases from 1080p to 1440p. Assuming Thraktor’s tests were controlled correctly (more on that below), the only explanation that I can think of is that Performance, Ultra Performance, and Quality modes already use somewhat different network architectures.
The way that a CNN works is independent of resolution because the weights are shared. That concept (parameter sharing) is one of the main advantage of CNNs, because it means there are fewer weights to train. Besides, from an image processing perspective, it would be a poor choice to use different weights for each pixel position, since the information at those pixel positions is different in each frame. So instead you train a filter, often a 3x3 or 5x5 pixel filter, that loops over each pixel in the frame.
In theory, since you have parameter sharing, you could use the same neural network for each mode. It would be a conscious decision on Nvidia’s part to train a different network for each mode to improve either cost or image quality. If this were true, Thraktor’s results suggest that Performance mode may use a deeper, more computationally expensive network than Quality mode.
This would make the names something of a misnomer, although it does make a certain amount of sense - for example, it makes intuitive sense that you would need a deeper network to reconstruct a “4K” looking image from 720p frames than from 1080p frames.
This doesn’t fully account for what happens when DRS and DLSS are combined. For example, which network would be used when scaling from 900p to 4K - performance or ultra performance? I am not sure.
I am hesitant to make any conclusions based on Thraktor’s tests alone. Taking the median frame time seems reasonable, but I am not sure if manually walking through a set path is controlled enough for the 0.3 ms margins we’re talking about. Plus, we haven’t seen any indication that the networks are different elsewhere. It’s an interesting possibility for the sake of discussion though.
(edited heavily to reorganize paragraphs and improve clarity)