F1 2019's DLSS support was "released prematurely" - Explains graphical issues

[WYP]

F1 2019's DLSS support was "released prematurely" - Explains graphical issues

A corrected implementation of DLSS is coming in a future F1 2019 patch.



Read more about F1 2019's DLSS support being "released prematurely".

 

[NeverBackDown]

Now that I think about this. This is a PR stunt. You cannot just simply forget that a large codebase is in your source code that gets pushed for an update. The game has been released already. They aren't adding all that much. But to forget a MAJOR feature codebase is still present is just one of the most unprofessional things I could think of that not only gets by every programmer and the Lead programmer, but Q&A as well AND then getting pushed.

I'm calling BS and Nvidia got mad at Codemasters and forced them to make a PR statement.

 

[grec]

The models that define the neural networks inference behaviour are made and provided by NVidia(And supplied to end users via NVidias drivers iirc, inference models are mostly just a big dataset of weights with no active code defining its operation), the developers hand them the source code for their games and NVidia runs them on at 64x resolution on their basically "mini supercomputer in a rack" machines and uses this as the training data(Target output) with the base game at standard resolutions as the models input to create/train the inference model. The more time this model gets to train the better it will be but there's significant diminishing returns over time and sometimes your back propagation can get a bit "stuck" if there's certain patterns or misconsiderations in the training data selection.

 

[NeverBackDown]

Not sure why you're explaining DLSS? That's not what the article is about...

 

[g0ggles1994]

This PR stunt has a ring of 'Daddy Nvidia gave us a telling off because AMD's looks better'

 

[grec]

My point is its not necessarily codemasters problem if they expect NVidia to have the DLSS models ready for launch. NVidia are obviously putting out some kind of early model to the public themselves if it's running and my guess is this was a late notification from NVidia that things wouldn't be ready for launch.

 

[WYP]

From my end, this seems to be more of a Codemasters problem then an Nvidia problem. Yes, the AI algorithm has an impact, but from my experience testing DLSS in games, it is more often the game updates that have more of an impact than driver updates. Monster Hunter World didn't even get a DLSS optimised driver.

I remember Metro Exodus getting a huge boost in DLSS quality when the game was patched, and little to no difference when the next DLSS driver released. Also note that Codemasters makes no mention of a new Nvidia driver, just a game update.

We will just have to see what happens when the game is next updated.

 

[AlienALX]

AMD have always been better at this kind of stuff. It's just a fact. Nvidia seem to spend more of their money elsewhere.

Tress FX was better than Hairworks and performance was much better too. Heck, it even ran better on Nvidia cards. Gamedoesn'twork was terrible and thus it doesn't surprise me at all that DLSS has so far been completely sub par, or that the AMD idea works very well.

It's a shame the adoption rates are not higher.

 

[grec]

From what I've read the training models are downloaded from NVidia servers via the driver once the application is detected, ie they're supplied via NVidia's drivers, but not with them, so they can be pushed on demand. Usually there's quite a few models though depending on resolution/settings/aspect ratio and how those are selected and applied will be game side, this definitely seems like a purely model related issue though whoevers fault it is for these being distributed and released(Personally I think both could have prevented this) as that water painting style effect is more or less what you'd expect from a model which hasn't be trained for long enough.