Ahead of this week's CES-timed GeForce On presentation, Nvidia hosted a hands-on event where the Digital Foundry team could go hands-on with nearly every aspect of the GPU maker's 2026 announcement roster. Here are our impressions of what we've seen and were impressed by.

DLSS 4.5 transformer model: Our pre-brief look at Nvidia's latest Super Resolution technology was limited to brief video snippets, but we've since looked at fuller real-time examples of the new "Preset M" transformer model enabled. Essentially, the transformer model has been trained on an even more elaborate visual data set to visually surpass the upsampling already available from last year's "Preset K" version of the transformer model.

In addition to improvements we'd previously seen in DLSS 4.5, like reduced camera-turn ghosting, particle effects and higher retained detail in background elements like foliage, we have now also seen improved handling of surface edges with specular highlights - a larger specular response - which we believe is thanks to the DLSS model now performing a pre-tone-mapping pass. This results in less compressed and clamped visual information in the accumulation model.

Since our demo, we've learned that the new Preset M and Preset L transformer models - which are now live for end users - have measured costs compared to DLSS 4.0's Preset K version. Nvidia's published stats offer a baseline measure of DLSS impact in "performance" mode, with a base 1080p image being upscaled to 2160p resolution. According to Nvidia's "ballpark" measurements, the K-to-M difference on an RTX 5090 is 20.7 percent, jumping from 0.87ms to 1.05ms.

That per-frame DLSS impact grows to as much as a 39 percent jump on the RTX 4070 Ti, though as a reminder, Ada Lovelace and Blackwell GPUs employ a floating-point 8 (FP8) process while running Preset M to reduce its per-frame processing time. While older RTX GPUs can also access Preset M, they do not have the same FP8 hardware support and thus do so at a higher processing length per frame. The K-to-M jump across a number of RTX 2000 and 3000-series GPUs is over 200 percent in time. As an example, on the RTX 3080 Ti, Preset K costs 2.06ms per frame, while Preset M costs 4.35ms per frame.

Preset L adds even more of a per-frame cost than Preset M across all RTX-branded GPUs, and it's advertised specifically to support DLSS's "ultra-performance" upscaling preset - and we plan to look further at the trade-off between performance and visual uplift we're getting from each new DLSS 4.5 preset. For the RTX 4000 and 5000 series, the small per-frame cost for Preset M's visual improvements immediately seems worthwhile.

DLSS 4.5 Dynamic Multi-Frame Generation: Our pre-release demos of Nvidia's new DMFG system come ahead of its wider public launch in "Spring 2026" as an RTX 5000-series GPU exclusive. They confirm a measurable-if-mild latency jump from the older 4x model to the newer 6x model - albeit with a negligible cost of roughly 4-6ms of input latency from the brief demos we captured off-screen. With this new "dynamic" handling of MFG, the latency count bounces up and down according to how high of a frame-generation multiplier is applied in a given moment, with less MFG leading to tighter input response.

Interestingly, Nvidia made The Outer Worlds 2 available as one of its DMFG demo games, albeit only in indoor spaces - as opposed to the outdoor, open-world scenes that we've confirmed continue to suffer from stuttering issues on PC. Still, it proved to be a compelling example of DMFG in action, as its bounces between 3x and 6x FG were nigh imperceptible.

We're also waiting to see the results of Nvidia's teased updates to the model that will reproject semi-transparent UI elements - maps, cursors, aiming reticules - more accurately into the current gameplay frame.

Pragmata with path tracing: Nvidia and Capcom gave us a surprise announcement that this RE Engine third-person action game will indeed enable its path-tracing mode on PC when it launches in April of this year. And while we're not surprised to see solid ray reconstruction and handsome reflections in the game's latest preview build, its path-tracing version does a much better job handling precise shadows than the aliased, saw-toothed shadows we have seen in other RE Engine titles.

The combined, resulting visual makeup looks particularly cool in the game's science fiction-inspired environs, marked by glossy plastics, glass windows and reflective metals. We're hopeful the demo's issues with "firefly" particle artefacts are resolved by the PC version's April launch date.

Other updates: Our look at the new RTX Remix Logic system does not go too far beyond what Nvidia publicly revealed in its GeForce On presentation, but we remain impressed by the amount of visual-impact modding that can now be achieved without users having any access to an origin game's source code - at least, as showcased in a few intense Half-Life 2 sequences. Screen and object distortion as triggered by "dangerous" enemies, sweeping weather patterns and even a post-nuclear treatment to environments and enemies alike all looked and felt impressive during our hands-on time.

We went eyes-on with G-Sync Pulsar, a long-in-development combination of monitor technologies that makes IPS panels present images in ways that far better resemble CRT screens - all while supporting variable refresh rate (VRR) presentation of frames. We'll have much more on this technology in the near future, but in our CES-adjacent demo, we did confirm Nvidia's interest in applying the same technology to OLED panels in the future.

Nvidia's bullishness about "small language model" AI is arguably most impressive in the form of locally rendered assistive systems, and we liked testing this out in 2023's Total War: Pharoah. By directing tapping into the massive database of instructions and numbers that impact moment-to-moment gameplay, and then combining that with a language model, strategy players can naturally type out questions about how the game functions and get real-time responses - all using nothing more than the built-in power of your GPU.

As Nvidia previously confirmed, this requires roughly 6GB of VRAM to run in TW:P's "experimental" test of the feature, while PUBG's own implementation of a real-time "buddy" CPU player as run on-device appears to cap out at 2GB of local VRAM.

In response to rumours, Nvidia confirmed that RTX 3060 GPUs aren't "returning" to production but rather have been in consistent production since their 2022 launch. When asked about its own GPU prices in the wake of rapidly evolving markets, Nvidia said it did not plan to change prices on its own "Founders Edition" GPUs - though we note that these are far less impactful to the general GPU supply in the wild compared to OEMs' own MSRP shifts based on supply and demand.