More info on XeSS
* The network does not require per game training, so like DLSS 2.0+ it's a plug and play solution for any game. Intel says it's also working on version 2 for XeSS.
* The XeSS neural network has been trained at 64 samples per pixel - that is 4 times higher pixel count than Nvidia uses for training, Intel says it's goal for XeSS is to be market leading and absolutely beat DLSS
* XeSS runs on any GPU as long as it has Shader model 6.4 or newer and you're using a version of windows that supports it - that means you need an up to date Windows 10 or Windows 11, for Nvidia it means any GPU from Pascal GTX1000 and higher is supported
I do not fully agree with your interpretation. What they said, in order you wrote above is as follows (quotes):
* "
There will be XeSS 2.0 at some point, XeSS 3.0 at some point. You know at some point maybe graphics just completely changes and it’s all neural networks." - that does NOT mean they're working on it. They just stated the obvious - things will change in the future and so their tech will have to be adjusted to it. In other part they said it's open source, hence they count on community changing and improving it.
* "
So, when NVIDIA says 16k images, I am assuming it translates to the number of samples it has inside a pixel. So from our standpoint, that’s what I can talk about. We train with reference images that have 64 samples per pixel." - Intel assumes here how it works on NVIDIA side of things, as they do not really know, because it's not a public information (sans marketing blabla). They also did not state anywhere they want to beat DLSS or anything like that?
* "
So, for DP4a, yes, SM 6.4 and beyond SM 6.6 for example supports DP4a and SM 6.6 also supports these packing intrinsics for extracting 8-bit data and packing 8-bit data. So we recommend SM 6.6." - working on SM 6.4 yes but for best performance SM 6.6 is recommended. For reference, SM 6.4 = Vega and RDNA 1 (AMD); Pascal (NVIDIA). RDNA 2 = 6.5 (AMD). Turing and Ampere = 6.6 (NVIDIA).
In addition, Intel said they will not support tensor cores (or AMD AI acceleration in the future) because they do not intend to optimise their solution for every single architecture. It's open source, devs can do it themselves if they want, but Intel believes it's fast enough with DP4a - which by itself is already a matrix acceleration, as people forget the whole GPU is a matrix accelerator as such, many times faster than CPU in such calculations. Tensor cores in most tests I've seen is only 10-20% faster than DP4a and this seems to be consistent with graphs Intel shown in regard to their hardware acceleration vs DP4a.
That also means that NVIDIA could most likely just let DLSS 2 run on DP4a with barely any speed loss and the whole tensor cores requirement is there just to boost sales of new GPUs. As is typical for them. Perhaps competition from Intel will force them to release it for older GPUs or even other brands as well.
