NVIDIA’s latest optimization update for its DLSS tech has been spotted by VideoCardz, as it invovles making transformer-based upscaling model leaner and better by cutting VRAM usage by up to 20%.
The report stated that the latest DLSS SDK 3.10.3.0 will make DLSS 4 “official” (Yeah, it has been a “beta release” ever since) as well as applying the aforementioned upgrade that drops memory footprint close to the older CNN-based versions of transformer models.
Apparently, before this update, transformer-based models cost almost 2x the VRAM of CNN models, which can be devastating to, you know, the infamous “8GB cards”. Welp, at least these GPUs can be stressed less when the update is being integrated into drivers and supported games.
While a percentage-based improvement figure may be impressive, let’s not forget that DLSS VRAM usage scales with resolution, where only around 80MB is saved at 4K, which scales approximately in a linear way when moving down to QHD and FHD resolutions. It can be something in 8K though, but then again, if you have the best GPU at your disposal, that 300MB saved for a 32GB pool might just be “placebo” at best or “negligible” at worst.
It is also worth noting that these VRAM optimizations apply specifically to DLSS upscaling and not frame generation (that’s another story where they cut 30% VRAM usage in DLSS 4).
All in all, this might be a small improvement during the early days for these transformer models, and the next leap might just be around the corner, just because it’s AI magic and it can just pop up the next day when you wake up.
