Friday, March 14th 2025
NVIDIA Details DLSS 4 Design: A Complete AI-Driven Rendering Technology
NVIDIA has published a research paper on DLSS version 4, its AI rendering technology for real-time graphics performance. The system integrates advancements in frame generation, ray reconstruction, and latency reduction. The flagship Multi-Frame Generation feature generates three additional frames for every native frame. The DLSS 4 later on brings the best looking frames to the user quickly to make is seem like a real rendering. At the core of DLSS 4 is a shift from convolutional neural networks to transformer models. These new AI architectures excel at capturing spatial-temporal dependencies, improving ray-traced affect quality by 30-50% according to NVIDIA's benchmarks. The technology processes each AI-generated frame in just 1 ms on RTX 5090 GPUs—significantly faster than the 3.25 ms required by DLSS 3. For competitive gaming, the new Reflex Frame Warp feature reduces input latency by up to 75%, achieving 14 ms in THE FINALS and under 3 ms in VALORANT, according to NVIDIA's own benchmarks.
DLSS 4's implementation leverages Blackwell-specific architecture capabilities, including FP8 tensor cores and fused CUDA kernels. The optimized pipeline incorporates vertical layer fusion and memory optimizations that keep computational overhead manageable despite using transformer models, which are twice as large as previous CNN implementations. This efficiency enables real-time performance even with the substantially more complex AI processing. The unified AI pipeline reduces manual tuning requirements for ray-traced effects, allowing studios to implement advanced path tracing across diverse hardware configurations. The design also addresses gaming challenges like interpolating fast-moving UI elements and particle effects and reducing artifacts in high-motion scenes. NVIDIA's hardware flip metering and Blackwell-induced display engine integration ensure precise frame pacing of newly generated frames for smooth, high-refresh-rate gaming, with accurate imagery.
To ensure DLSS works as intended and that the neural networks produce quality results, NVIDIA has used a secret weapon: a dedicated supercomputer that has been continuously improving DLSS for the past six years. The supercomputer's primary task involves analyzing failures in DLSS performance, such as ghosting, flickering, or blurriness across hundreds of games. When issues are identified, the system augments its training data sets with new examples of optimal graphics and challenging scenarios that DLSS needs to address. That way, DLSS learns what games look like and generates realistic frames like a game engine would, without any artifacts.
Source:
NVIDIA Research
DLSS 4's implementation leverages Blackwell-specific architecture capabilities, including FP8 tensor cores and fused CUDA kernels. The optimized pipeline incorporates vertical layer fusion and memory optimizations that keep computational overhead manageable despite using transformer models, which are twice as large as previous CNN implementations. This efficiency enables real-time performance even with the substantially more complex AI processing. The unified AI pipeline reduces manual tuning requirements for ray-traced effects, allowing studios to implement advanced path tracing across diverse hardware configurations. The design also addresses gaming challenges like interpolating fast-moving UI elements and particle effects and reducing artifacts in high-motion scenes. NVIDIA's hardware flip metering and Blackwell-induced display engine integration ensure precise frame pacing of newly generated frames for smooth, high-refresh-rate gaming, with accurate imagery.
62 Comments on NVIDIA Details DLSS 4 Design: A Complete AI-Driven Rendering Technology
You said AMD and Intel were Developing this stuff before Nvidia Released it.
but people read... AMD and Intel were Developing this stuff before Nvidia Developed it.
:shadedshu:
As an example AMD has had Matrix cores since CDNA1, but that's useless to the regular consumer market and only now has reached a consumer product. Even on the enterprise market it's kinda moot since their software stack can't make proper use of those.
As for higher prices, you can blame the death of Dennard scaling, the growth of AI and crypto, the maturation and mass adoption of CUDA, ever-slowing ever-increasingly costly fab nodes, the higher proportions of sales now going to deep-pocketed corporate buyers instead of retail sales, the pandemic changing habits, EVs and other products requiring more chips, higher interest rates, more expensive shipping, inflation in general, etc, etc.
But if you are implying that these companies did not care about maximizing profits back in 2005 or 2010 or 2015 but now they do care and this is why the prices are so high now but were not in the past, I have a massive AI-enhanced bridge to sell to you. :kookoo:I took a look at Anandtech's review of the HD 7970 back in 2012 and it achieves 3.79 Tflops (of what, FP32, not specified at all) vs the 48.7 Tflops (of FP32) that the 9070 XT can achieve. Pretty nice improvement in performance considering that Moore's law has been dying the whole time.
Also to correct myself: the computer I meant to compare was the Blue Gene/Q. The /L is the one from 2004, which in its highest configuration was just about as powerful as a single RTX 4090 (actually a little slower)
Even FSR4 is a hybrid CNN/Transformer model according to AMD.
The fight for technological superiority must go on :p
DLSS is the best thing since baked bread. I'll gladly sacrifice the minuscule loss of quality, (that most people hardly ever notice when actually playing the game, not pore over comparison screenshots or vids) for the massive fps gains that let me play latest games in 4K, often maxed out and with RT to boot.
Yeah, sure, it'd be great if you could play them in native res. But the fact that a GPU that let’s you do that doesn't even exist should tell you something about the state of affairs.
Being first to market has tons of benefits in any kind of business.
Everything has a 1.0 release.
DE allows people to play the pre-alpha version of their upcoming game Soulframe, but its such a nothing product at the moment that I would not recommend it to anyone, there is virtually nothing to play because obviously Pre-Alpha, I know they want to gather feedback but I would not have exposed this ti the public at this time...and its been excisable for like 2 years now.....Re-read the conversation, this is not the type of thing you would ask if you understand what it was about, sorry.
- Increase the number of transistors in the GPU chip.
- Develop new algorithmic techniques (AI is also one of them).
So, what’s the result? More transistors mean exponentially higher costs, making it unaffordable for many users. On the other hand, improving architecture and optimizing algorithms only requires time investment, without significantly increasing manufacturing costs.But I shouldn't have to enable DLSS/FSR for something like Monster Hunter Wilds to get 60fps and higher at 1440p! The visual in the majority of modern games that require DLSS/FSR just don't justify the absurd resource cost they ask for, that's all.
Do you think Monster Hunter Wilds would be better optimized had upscaling not implemented?
Also you are not forced to enable Upscaling if 30FPS is all you need anywaysYup, as tech lover I'm always excited for the break-through techs that could improve my gaming experience further, whether it's from a new CPU, GPU or a Monitor (4K 480hz OLED screen would be sick).
I think you might not be completely correct on needing "only" the time investment. Time investment is cost. If you think about all the things GPU hardware vendors do - especially Nvidia with their software focus - then they spend a lot of time and money in that area.