NVIDIA Editor's Day: Blackwell, RTX 50, DLSS 4 and Multi Frame Generation
Today, finally, the embargo on information revealed during the Editor's Day from NVIDIA, held on January 8th within the framework of the CES 2025 in Las Vegas, United States, the largest technology convention on the continent. This event touched on a wide variety of topics, highlighting the new architecture that powers graphics cards NVIDIA GeForce RTX 50, which incorporate innovative technologies such as Neural Rendering, DLSS-4, DLSS Multi Frame Generation, among others. Due to the complexity and amount of detail addressed, this will be an extensive analysis.
It is important to note that all information presented here, including percentages, data and technical specifications, has been provided directly by NVIDIA.
Table of Contents
NVIDIA Blackwell Architecture: Innovation Beyond Silicon
It's been a long time since Turing was launched and some Jerusalemite geniuses launched their prophecy that RTX had no future and that these products were dead on arrival.
Since the launch of the first RTX 20, NVIDIA began to revolutionize the graphics card market by introducing technologies such as Ray Tracing (ray tracing) and AI-powered superscaling, known as DLSS Super ResolutionThese innovations were improved with the launch of the RTX 30 y RTX 40, standing out with the introduction of DLSS Frame Generation in the Ada Lovelace series.
With the new architecture BlackwellNVIDIA is strengthening its focus on once again changing the paradigm of traditional rendering, introducing key technologies that will work especially well with the new cards. GeForce RTX 50. Some of the major innovations include DLSS Multi Frame Generation, NVIDIA Reflex 2, Neural Rendering and RTX Neural Shaders, which optimize performance and visual quality in real time.
Innovations in Blackwell SMs (Streaming Multiprocessor, SER 2X, GGDR7, RT and Tensor Cores)
One of the main objectives of architecture Blackwell, according to NVIDIA, is to overcome the diminishing returns and limitations imposed by the famous Moore's LawTo achieve this goal, NVIDIA has partnered with TSMC 4NP to utilize a custom manufacturing process that integrates enhanced components such as:
- New SM cores (Streaming Multiprocessor), optimized for complex rendering tasks.
- Fifth Generation Tensor Cores, designed to accelerate AI and advanced graphics applications.
- A specific processor for efficient management of tasks AI (AI Management Processor).
- GDDR7 memory offering greater bandwidth and better energy efficiency.
These components allow the new generation of graphics cards to offer improved performance, optimizing each of the processes within the rendering and artificial intelligence pipeline.
Neural Shaders: A Breakthrough in SMs
On graphics cards RTX 40 (Ada Lovelace) and earlier models, NVIDIA used traditional shaders, distributed in separate clusters that handled the integer and floating-point instructions.
With architecture BlackwellThe new SM allows the cores shaders execute both integer and floating-point instructions within a single cluster, working in conjunction with the new cores Fifth generation tensorThis improves processing efficiency and reduces the need for additional resources.
Improved Shader Execution Reordering (SER 2X)
Technology Shader Execution Reordering (SER), introduced with the RTX 40, allowed a significant reduction in divergence in shader executions, optimizing performance in scenes and jobs that use Ray Tracing.
With Blackwell and cards GeForce RTX 50, this technology has been further improved, increasing its efficiency to the double compared to Ada Lovelace (RTX 40). In addition, it is now able to more intelligently evaluate which cores (Tensor or Shader) to use for each specific task, avoiding bottlenecks and optimizing resource usage, maintaining consistency between data type divergence.
New GDDR7 memories (higher bandwidth, lower voltage and greater efficiency)
The new memoirs GDDR7 they use Press Amplitude Modulation 3 (PAM3), a signal modulation technology that allows the use of different voltage levels, thus improving energy efficiency and reaching new levels of data transfer. According to NVIDIA, graphics cards RTX 50 equipped with Blackwell will reach a bandwidth up to 30 GB/s in its most powerful model, which doubles the efficiency of the memories GDDR6 y GDDR6X, which were introduced in 2017 and 2022.
Fourth Generation RT Cores: Evolution in Ray Tracing
The architecture Ada Lovelace (RTX 40) introduced the third generation of RT (Ray Tracing) cores, which include specialized engines such as the Box Intersection Engine (engine for box intersection) and the Opacity Micromap Engine (engine for opacity micromaps). With Blackwell, the fourth generation of these cores improves performance and optimizes the triangle intersection engines (Triangle Intersection Engine). This triangle engine is now split into two components:
- Triangle Cluster Intersection Engine: Optimizes the intersection of groups of triangles to improve performance in complex scenes.
- Triangle Cluster Decompression Engine: decompresses groups of triangles, facilitating a more efficient representation of complex geometries.
In addition, the new component Linear Swept Spheres (LSS) becomes a key element for new technologies such as RTX Neural Faces, which improve the precision of ray tracing, achieving a higher level of detail without penalizing performance. Also, the concept of “Mega Geometry” allows developers to generate up to 100 times more triangles in scenes, creating more realistic images without affecting rendering speed.
Also, a new key component is introduced:
- Linear Swept Spheres (LSS): linear scanning spheres, essential to enable technologies such as RTX Neural Faces, which optimize ray tracing to render human faces more realistically.
With these improvements, Blackwell brings to life “Mega Geometry”, a technology that allows processing up to 100 times more triangles in ray tracing, thanks to the optimization of the BVH (Boundary Volume Hierarchy). This allows for generating images with higher levels of detail (LOD) without affecting performance.
In summary: More triangles, more realism and without significant performance penalties.
Fifth generation of Tensor Cores: FP4 and advances in artificial intelligence
The 5th generation Tensor Cores in architecture Blackwell include a new engine for 4-bit floating point (FP4), which, according to NVIDIA, significantly improves performance in tasks related to artificial intelligence, offering up to 32 times more performance compared to video cards GeForce GTX 10 Series (Pascal).
This improvement facilitates the use of large language models (LLM) in applications and games, optimizing the performance of technologies such as DLSS and Frame Generation.
In conjunction with Tensor cores, the AI Management Processor (AMP) is another key component of architecture BlackwellThis processor is designed to manage tasks related to artificial intelligence more efficiently, helping to optimize the distribution of processing resources and improve latency and frame pacing.
With AMP, GeForce RTX 50 They can more effectively handle AI jobs, contributing to better performance on complex tasks that require advanced processing.
Blackwell on Laptops: Innovations in Efficiency and Performance
Technology Max Q NVIDIA’s GPU, designed to optimize performance and energy efficiency in RTX laptops, receives significant improvements with Blackwell:
- Advanced Power Gating: Quickly disable unused sections of the GPU to reduce power consumption.
- Low Latency Sleep: Allows the GPU to enter an ultra-low “hibernation” mode when not in use, saving battery even during light tasks.
- Accelerated Frequency Switching: Adjusts clock speeds in milliseconds to balance power and performance based on workload.
Finally, the adoption of GDDR7 and its more advanced voltage controls contribute to improving overall efficiency and extending the lifespan of Max-Q certified laptops, setting a new standard in portability and performance.
Outputs and video at Blackwell: DisplayPort 2.1 UHBR20 and more
The NVIDIA Blackwell architecture introduces significant changes to the video output engine, highlighting the arrival of DisplayPort 2.1 UHBR20. This standard offers a speed of 20 Gbps per line, far exceeding the 8.1 Gbps of DisplayPort 1.4a HBR3 of Ada Lovelace (RTX 40). With a total bandwidth of 80 Gbps (four lanes), DisplayPort 2.1 allows resolutions up to:
- 8K UHD (7680 × 4320) at 165 Hz.
- 4K UHD at 480 Hz.
In addition, the High Speed HW Flip Metering, a key component for new technology DLSS Multi Frame Generation (multi-frame generation).
NVIDIA also introduces advances in video encoding and decoding:
- 9th generation of encoder.
- 6th generation decoder.
These new components support formats such as AV1 UHQ, 2x H.264 Decode, MV-HEVC, and 4:2:2 Encode and Decode, making it ideal for applications of professional video production high quality.
RTX Neural Rendering: The future of gaming rendering
El RTX Neural Rendering is one of Blackwell's most innovative technologies and promises to revolutionize graphics in video games (it was the first talk during Editor's Day). According to NVIDIA, this architecture is designed to take advantage of neural networks through engines and functions optimized specifically for the neural rendering.
“The NVIDIA RTX Blackwell architecture is built and optimized for neural rendering. It has massive processing power, with new engines and features specifically designed to accelerate the next generation of neural rendering.”
NVIDIA Blog post
RTX Neural Shaders: Neural Networks in Programmable Shaders
During the NVIDIA Keynote (in which xanxogaming estuvo present), Jensen Huang, CEO of NVIDIA, explained the integration of neural networks into programmable shaders:
“And now we have the ability to intermix AI workloads with computer graphics workloads, and one of the amazing things about this generation is that the programmable shader is now also capable of processing neural networks. The shader is able to carry these neural networks, and as a result, we have invented neural texture compression and neural material shading.
As a result, we get these incredibly beautiful images that are only possible because we use AI to learn the texture, learn the compression algorithm, and as a result, get extraordinary results.”
The Blackwell RTX Neural Shaders They use artificial intelligence models to improve visual quality in real time. To put it simply, a neural network is an artificial intelligence model that mimics the functioning of the human brain. Examples such as DLSS 2 and DLSS 3 demonstrate how training neural networks can improve image fidelity as they are optimized.
with this technology (RTX Neural Shaders), new features are enabled for video games:
- Cache of Radiance.
- Neural texture compression.
- Neural shading of materials.
- Fields of Radiance and more.
These innovations will be available to developers through the NVIDIA RTX Kit, which includes tools such as RTX Neural Shaders SDK y RTX Neural Texture Compression SDK at the end of the month.
Note to developers: How to train neural networks with RTX
El RTX Neural Shaders SDK allows training game data and shaders on RTX-equipped PCs, using Tensor Cores to accelerate real-time neural representations. This process compares neural data with traditional data, refining it over multiple cycles.
To simplify training, developers can use Hose, a language that breaks complex functions into manageable chunks, making processing easier and improving graphics optimization.
Three Technologies to Rule Them All: Compression, Materials, and Glow Cache
“One Ring to rule them all” – JRR Tolkien
RTX Neural Texture Compression is one of the most important innovations within the new programmable shaders with neural networks. This technology uses artificial intelligence to compress thousands of textures into a model, achieving this in less than a minute. According to NVIDIA, this can save up to 7 times storage space in VRAM or RAM compared to conventional textures used in current games. In addition, this compression process is the basis for other technologies within the Neural Rendering.
RTX Neural Materials, meanwhile, takes this idea a step further by compressing complex materials into programmable code. This allows for high-fidelity real-time rendering of materials such as porcelain or silk, in cinematic quality and at a frame rate suitable for video games.
Finally, RTX Neural Radiance Cache (NRC) addresses real-time lighting using neural networks. This model is trained to estimate indirect lighting more quickly and accurately, reducing the resources needed for techniques such as path tracing.
NRC It traces one to two rays and saves them (storage and training) and with that information infers the unlimited amount of rays for a scene. In simple terms, this is a more efficient and intelligent way to produce an image or real-time rendering in Ray Tracing and Path Tracing, using less resources (less rays traced).
In the demos that were shown to us at Editors Days, it showed how to apply RTX Neural Radiance Cache (On vs Off) generated a scene (several seconds video) where the shadows of the leaves were more realistic. These details with greater fidelity are based on the rays traced indirectly within the scene (using NRC).
The future is now: RTX Neural Faces
NVIDIA hasn't stopped at textures and lighting; now it's taking neural networks to the next level with RTX Neural FacesThis technology uses trained models to capture and reproduce human faces with a high level of detail. Variables such as lighting, expressions and occlusion are considered to create ultra-realistic characters, perfect for NPCs in games or to increase immersion in interactive scenes.
In the pre-training process, the model can use both real photographs and AI-generated images, optimizing them by NVIDIA TensorRT for real-time applications. Developers will be able to apply these models in game engines to enhance facial realism or generate unique characters with next-generation visual fidelity.
Skin and hair: The next level
As mentioned earlier, Mega Geometry It is one of the bases of the new cards GeForce RTX 50. Within this context, technology LSS (Linear-Swept Spheres) plays a crucial role in creating realistic elements such as the hair and skin of the characters.
Traditionally, ray tracing for hair required generating multiple triangles per strand, which was resource intensive.
With LSS, NVIDIA uses spheres connected by lines to represent hair, dramatically reducing geometric complexity. According to NVIDIA, this technique consumes up to three times less information, optimizing VRAM usage and improving frames per second (part of RTX Character Rendering SDK).
The highlight of these technologies was the demo titled Zorah, which shows the potential of Neural Rendering combined with Mega Geometry to create a new level of visual gaming experience.
We close with one of the most intriguing surprises of the Keynote at CES 2025, where Jensen Huang hinted at the arrival of DLSS-4While details were scarce during the presentation, all indications are that this new iteration will take NVIDIA's super-resolution technology to a new level.
DLSS 4 – Reinventing upscaling and frame generation
If you are familiar with DLSS (Deep Learning Super Sampling), you know that this technology revolutionized super scaling in its version 3.0 through artificial intelligence, Tensor Cores and the introduction of DLSS Frame Generation, driven by the Optical Flow AcceleratorNow, with DLSS 4, NVIDIA is taking things to the next level, promising improvements for All existing GeForce RTX cards and some card exclusivity Blackwell GeForce RTX 50.
Key new features include:
- DLSS Frame Generation: Better performance and lower memory usage (exclusive to RTX 50 and RTX 40).
- DLSS Ray Reconstruction: More stable and detailed lighting in Ray Tracing (available for all GeForce RTX models).
- DLSS Super Resolution: Sharper images in motion, currently in Beta (all GeForce RTX models).
- Deep Learning Anti-Aliasing (DLAA): Better detail and stability in motion, also in Beta (all GeForce RTX systems).
These improvements are made possible by a new transformative DLSS model.
CNNs, Transformers and DLSS: From theory to practice
The CNNs (Convolutional Neural Networks) They were pioneers in AI, applying it to image classification and object detection. This model has been the basis of the classic DLSS (DLSS CNN).
However, the arrival of the Transformers (robots in disguise…), and more specifically the Vision Transformer (ViT), marked a revolution in image processing, introducing the concept of “tokens” and allowing more complex data to be analyzed with greater precision.
Now NVIDIA has integrated the Transformer model into DLSS (DLSS Transformer Model), and the results according to NVIDIA are remarkable:
- Double parameters.
- Four times more computing power.
- Better image quality compared to CNN-based DLSS.
The result? A more fluid, detailed and precise visual experience, which takes shape with DLSS Ray Reconstruction, as could be seen in Alan wake 2.
DLSS Multi Frame Generation – Multiplying frames
The generation of additional frames is not new; in DLSS 3 Frame Generation, the AI created one generated frame for every rendered frame. However, this method could introduce bottlenecks by handling more than one generated frame (e.g. two AI-generated frames) for every input render to the GPU.
DLSS 4 Multi Frame Generation breaks this limitation, generating up to three additional frames for each rendered frame.
How does he do it?
- El Optical Flow Accelerator of the RTX 40 has been replaced by the AI Optical Flow, a more advanced AI model.
- To ensure a stable frame rate, NVIDIA implemented the Flip Metering, a hardware component integrated into the Blackwell Display Engine (mentioned above).
These innovations make the process a 40% faster, consume a 30% less VRAM and requires only one frame as input to generate new ones.
Combining DLSS 4 and GeForce RTX 50 allows to generate 15 out of 16 pixels, ensuring high quality images with low latency thanks to reflex 2.
In addition, NVIDIA claims that performance can reach up to 8 times more (in x4 mode) compared to rendering without DLSS (conventional raster).
Ready-made games and customization options (NVIDIA App update)
NVIDIA promises that 75 games will support DLSS 4 at launch of the GeForce RTX 50. For titles without initial support, users will be able to manually customize the use of frame generation, Multi Frame Generation y Super Resolution, even in games without native DLAA.
NVIDIA Reflex 2 – Lower Latency and Coming Soon
Since its introduction in 2020, NVIDIA SLR revolutionized esports by drastically reducing latency in competitive games. Over time, its adoption expanded to AAA games as well, especially in light of the increased latency brought on by technologies such as DLSS Frame Generation. Reflex optimizes synchronization between CPU and GPU, ensuring that player actions (such as a mouse click) appear on screen in the shortest possible time.
Now with SLR 2, NVIDIA promises to reduce PC latency by up to 75%.
As it does?
It combines the low latency of the original Reflex with a new technology called Frame Warp, which analyzes the last mouse position to dynamically adjust the frame when rendering.
What is Frame Warp?
In simple words: while the GPU renders a frame, the CPU calculates the camera position of the next frame based on the most recent mouse movement. This is where Frame Warp:
- The CPU takes the new position of the camera.
- The GPU warps the frame that is already being rendered, adjusting it to this new position just before sending it to the screen.
- The result: an updated camera perspective that instantly reflects the player's movement.
The benefit? Drastically reduced latency, making every mouse movement more precise. This is especially crucial for targeting moving enemies or reacting to unexpected threats, resulting in an almost organic connection between player and game.
Inpainting: Repairing the holes
The marketing process includesseveral phases that are reflected below: Frame Warp involves shifting pixels in real time, which can create small “holes” in certain regions of the image, especially during fast mouse movements. However, NVIDIA has a solution:
- Predictive inpainting: Using data from previous frames (including camera, color and depth information), Reflex 2 accurately fills in those gaps, ensuring a smooth image free of visual artifacts.
In practical terms, this approach allows latency to be reduced by up to a whole full frame, lowering it to a Additional 50% compared to Reflex 1.
Reflex 2: Ready for the most popular games
NVIDIA Reflex 2 is coming soon to titles like The Finals y Valorant, where every millisecond can be the difference between winning or losing.
With a combination of innovations like Frame Warp and Inpainting, Reflex 2 doesn’t just improve latency; it redefines what it means to “be connected” to your game.
NVIDIA GeForce RTX 50 Series Graphics Cards – RTX 5090, 5080, 5070 Ti, and 5070
NVIDIA GeForce RTX 5090
Graphics Card | GeForce RTX 5090 |
NVIDIA CUDA Colors | 21760 |
Shader Colors | Blackwell |
Tensor Cores (AI) | 5th Generation 3352 AI TOPS |
Ray Tracing Colors | 4th Generation 318 TFLOPS |
Boost Clock (GHz) | 2.41 |
Base Clock (GHz) | 2.01 |
Standard Memory Config | 32 GB GDDR7 |
Memory Interface Width | 512 bits |
Display Support | 4K at 480 Hz or 8K at 120 Hz with DSC |
NVIDIA GeForce RTX 5080
Graphics Card | GeForce RTX 5080 |
NVIDIA CUDA Colors | 10752 |
Shader Cores (AI) | Blackwell |
Tensor Cores | 5th Generation 1801 To TOPS |
Ray Tracing Colors | 4th Generation 171 TFLOPS |
Boost Clock (GHz) | 2.62 |
Base Clock (GHz) | 2.30 |
Standard Memory Config | 16 GB GDDR7 |
Memory Interface Width | 256-bit |
Display Support | 4K at 480 Hz or 8K at 120 Hz with DSC |
NVIDIA GeForce RTX 5070 Ti (AIB model available only)
Graphics Card | GeForce RTX 5070 Ti |
NVIDIA CUDA Colors | 8960 |
Shader Cores (AI) | Blackwell |
Tensor Cores | 5th Generation 1406 To TOPS |
Ray Tracing Colors | 4th Generation 133 TFLOPS |
Boost Clock (GHz) | 2.475 |
Base Clock (GHz) | 2.30 |
Standard Memory Config | 16 GB GDDR7 |
Memory Interface Width | 256-bit |
Display Support | 4K at 480 Hz or 8K at 120 Hz with DSC |
NVIDIA GeForce RTX 5070
Graphics Card | GeForce RTX 5070 Ti |
NVIDIA CUDA Colors | 6144 |
Shader Cores (AI) | Blackwell |
Tensor Cores | 5th Generation 988 To TOPS |
Ray Tracing Colors | 4th Generation 94 TFLOPS |
Boost Clock (GHz) | 2.51 |
Base Clock (GHz) | 2.165 |
Standard Memory Config | 12 GB GDDR7 |
Memory Interface Width | 192-bit |
Display Support | 4K at 480 Hz or 8K at 120 Hz with DSC |
NVIDIA GeForce RTX 50 Laptops (Slides)
Final note – Xanxo “Rambling”
Well, the lifting of the embargo on the new GeForce RTX 50 series graphics cards (Blackwell) is approaching and This round of reviews is going to be brutal. Not only is there only that Analyze the generation gap in conventional rasterization, if not now too DLSS 3/4, as well as the comparison between Frame Generation y Multi Frame Generation.
I hope they give some air to all the media, because it will be honestly a pretty “brutal” round of reviews. (excuse the redundancy, I am aware of it) and sorry there wasn't time to add the news for content creators with the new Blackwell architecture
This was my first Editor's Day from NVIDIA (Denny's is evil for delaying lunch).
Until the embargo is lifted!
