Honestly agree here.
The Tegra X1 never really seemingly encountered any real fill-rate problems. Not at least before Memory or CPU became a problem.
Even when doing sophisticated upscaling techniques like Monolithsoft's TSR in Xenoblade 3, the game gets memory limited when alpha-effects are in place which makes sense considering the speed of the LPDDR.
Heck, the main symptom of Fill-Rate may be how low the internal resolution for Xenoblade 2/3 has to be for it to work as it does with how graphics-pushing the X engine can be at times in regards to scale and quality of things like material properties and geometric complexity for environments.
Which DLSS already covers that for Drake, but something else to point out really is that do we have the specific data on the ROPs in Drake
@oldpuck @Thraktor?
Or is that just assuming off the ROPs/GPC in Ampere/Lovelace.
I'm asking this as no Ampere card has had less than 32ROPs outside of the Binnned down Orin variations (Orin 32, the Orin NX duo, the lattermost breaking down to only 8ROPs).
Even the 10SM A2 card has 32ROPs still and that is a derivative of GA107.
The versatile, entry-level GPU that brings NVIDIA AI to any server.
www.nvidia.com
It just feels weird to me that they'd not up the ROPs to at least 24 like the Orin 32GB (14SMs) and purposely limit it to 16ROPs. Especially as Drake is running some variation of customization to the architecture beyond what was done with Orin.