During CES 2025, NVIDIA unveiled the GPU based on the Blackwell architecture and showcased the performance and features of NVIDIA RTX AI technology at its Editor’s Day event. Subsequently, NVIDIA held a further communication sharing session in Shenzhen, detailing the Blackwell architecture GPU and its functionalities. So, what other aspects are worth our in-depth exploration?
The NVIDIA GeForce RTX 50 series GPU is designed based on NVIDIA Blackwell, featuring a plethora of cutting-edge technologies.
The complete Blackwell core is equipped with the fifth-generation Tensor Core, fourth-generation RT Core, GDDR7 memory, and 360 RT TFLOPS, a staggering performance benchmark. It also supports FP4, RTX Mega Geometry, and boasts exceptional energy efficiency.
The SM units under the Blackwell architecture adopt Neural Shaders, which aim to embed miniature AI networks into programmable shaders to achieve cinematic quality material and lighting effects. By optimizing the graphics rendering process through neural rendering, Neural Shaders significantly enhance performance, image quality, and interactivity.
Shader Execution Reordering (SER) can reorder the GPU’s tasks to ensure that the SM runs with less code sent. In the example above, the Neural Shader is generating a blend of traditional shader code and neural rendering code; SER can recombine these different rendering workflows, significantly improving the efficiency of Tensor Core and Shader Core.
The GPU based on the Blackwell architecture is the first product to adopt GDDR7 memory, providing double the speed of GDDR6 while consuming only half the power. This is because GDDR6 and GDDR6X use Pam4 signaling with four voltage levels, while GDDR7 uses Pam3 signaling with only three voltage levels, allowing for a larger voltage eye and enabling higher data transfer rates at lower voltages, achieving 30Gbps with a higher energy efficiency ratio.
Beyond neural rendering technology, the process of achieving realistic effects demands higher geometric detail. The fourth-generation RT Core under the Blackwell architecture has been significantly optimized, introducing RTX Mega Geometry technology. This technology dramatically reduces the time required for acceleration structure construction, with reductions of up to 10 to 100 times, greatly increasing the number of geometries in each scene. The core upgrades include the Triangle Cluster Intersection Engine, Triangle Cluster Decompression Engine, and Linear Swept Spheres. We will delve into more details about these technologies in the following content.
Compared to the Ada Lovelace architecture, the introduction of RTX Mega Geometry in the Blackwell architecture has doubled the speed of triangle intersections, and the geometry compression on Blackwell has significantly reduced the space and bandwidth requirements of the acceleration structures.
According to NVIDIA’s provided breakdown diagram, we can see that the PCB of the NVIDIA GeForce RTX 50 series is located in the center of the graphics card, adopting a short board design. Meanwhile, the heat pipes distribute the heat generated by the PCB to the cooling fans on both sides for superior cooling performance.
The comparison of certain features among the NVIDIA GeForce RTX 30, RTX 40, and RTX 50 series shows that the upgrades of the RTX 50 series are comprehensive, including shaders (support for neural rendering), RT Core (support for RTX Mega Geometry), Tensor Core (support for FP4), DLSS (support for DLSS 4), encoders/decoders (iterative upgrades and increased quantity), and memory (using GDDR7), among others.
Currently, four products from the NVIDIA GeForce RTX 50 series have announced their prices, with the RTX 5090 D and RTX 5080 officially available on January 30, and the RTX 5070 Ti and RTX 5070 set to be released in February. Furthermore, considering NVIDIA will release SUPER and Ti SUPER related products later, there is significant pricing flexibility between the RTX 5090 D and RTX 5080.
In addition to the desktop version, the NVIDIA GeForce RTX 50 series will also be used in laptops. NVIDIA states that the RTX 50 series will have far better energy efficiency than the RTX 40 series, even the RTX 5070 laptop will outperform the RTX 4090 laptop, which enhances productivity and AIGC performance. Additionally, it will further help laptops increase battery life and reduce size, allowing OEM manufacturers to create portable yet high-performance laptop products.
The approximate specifications and related prices for laptops based on the NVIDIA GeForce RTX 50 laptop GPU show that even the RTX 5070 laptop has 798 AI TOPS and 8GB of memory.
RTX Neural Materials technology compresses the complex shader code of offline materials using AI technology. These materials typically consist of multiple layers, such as those used for rendering ceramics, silk, and other scenes. The processing speed of neural materials has increased fivefold compared to traditional methods, making it possible to achieve movie-level image quality at game-level frame rates.
RTX Neural Radiance Cache (NRC) is a neural network technology trained using real-time game data that can estimate indirect lighting in game scenes with greater accuracy and efficiency. NRC tracks a limited number of 1 to 2 light rays and stores this information in the radiance cache, inferring the paths and reflection effects of countless rays, thus reproducing indirect lighting in game scenes more accurately. This processing method not only enhances the performance of path tracing technology in indirect lighting but also reduces the number of rays that need to be tracked, thereby improving overall performance. Notably, NRC technology has now been integrated into the RTX global illumination SDK.
Additionally, NVIDIA demonstrated how its engineers developed a solution called RTX Neural Faces using the RTX neural shader architecture, which significantly enhances the realism of character faces in games using AI technology. Unlike traditional rendering techniques, the RTX Neural Faces solution creates more natural facial expressions by utilizing generative AI models after obtaining basic rasterized facial images and 3D pose data, effectively enhancing GPU performance. Rendering human faces in real-time graphics processing is a challenging task, as humans are extremely sensitive to facial features of peers, and any slight deviation can be detected, leading to the so-called “uncanny valley effect”. RTX Neural Faces offers an innovative approach to optimize facial quality through generative AI. Unlike direct rendering, RTX Neural Faces only requires simple rasterized faces and 3D pose data as a foundation to infer natural facial expressions in real-time through generative AI models. Previously, this model had been trained on thousands of offline data, covering various angles, lighting, emotions, and occlusion conditions.
Moreover, RTX Neural Faces complements the RTX Character Rendering SDK, which is specifically designed to enhance the realism of game character hair and skin. Achieving realistic effects for game character hair and skin is undoubtedly a challenging task. However, even with the most advanced technology currently available, traditional methods still require assigning 30 triangles for each strand of hair, while an entire hairstyle can require up to 4 million triangles to construct. This approach is not only costly but also slow in rendering. To address this issue, the GeForce RTX 50 series introduces a technology called Linear Swept Spheres (LSS). LSS technology reduces the number of rendered strands of hair and replaces triangles with spheres, allowing for a more accurate representation of hair shape. This innovation makes ray tracing for hair feasible while occupying less memory.
It is worth noting that NVIDIA has launched a highly efficient organization acceleration structure called RTX Mega Geometry, enabling developers to generate up to a hundred times the number of ray-traced triangles. Additionally, with the help of NVIDIA Opacity Micro-Maps technology, developers can encode the transparency of complex materials more accurately, ensuring near-realistic lighting effects in complex scenes. Through RTX dynamic lighting technology, integrations can achieve precise lighting effects; while the latest ReSTIR path tracing algorithm focuses on the main light paths, optimizing the allocation of computational resources. The RTX global illumination technology is AI-driven, effectively reducing the computational load required for ray-traced reflections.
Based on the Dragon sample showcased on-site, we can see that the number of triangles rendered is countless. RTX Mega Geometry can intelligently batch-generate triangle clusters on the GPU, relieving the CPU’s burden and enhancing the performance and image quality of ray-traced scenes. Additionally, NVIDIA announced that RTX Mega Geometry will soon be integrated into the NvRTX branch of the Unreal Engine to assist the Unreal Engine Nanite geometry system in efficiently completing ray tracing projects.
According to statistics provided by NVIDIA, currently, over 540 games and applications support Deep Learning Super Sampling (DLSS) technology. Among the 20 best games selected globally in 2024, 15 games have already implemented DLSS technology. Moreover, over 80% of NVIDIA users have utilized DLSS technology, with a cumulative usage time exceeding 3 billion hours.
NVIDIA’s DLSS technology, powered by GeForce RTX Tensor Core, has evolved into a mature solution through the latest version iterations. This technology not only enhances game frame rates but also provides clearer and higher quality image output. DLSS 4 introduces one of the significant innovations: Multi Frame Generation technology, specifically designed for GeForce RTX 50 series GPUs. Therefore, in the foreseeable future, only desktop and laptop computers equipped with GeForce RTX 50 series will enjoy the advantages brought by this technology. NVIDIA claims that at the launch of DLSS 4, 75 games and applications will initially support DLSS 4 Multi Frame Generation technology.
DLSS Multi Frame Generation technology generates three additional frames of images for each frame using AI technology based on traditional rendering methods, working in conjunction with other DLSS technology components.
Since DLSS technology allows games to render at lower resolutions and output at full resolution with high quality through algorithms, it has achieved performance improvements equivalent to eight times that of traditional rendering technology. This enables the NVIDIA GeForce RTX 5090 to run games smoothly at 4K resolution and 240 frames per second with the highest ray tracing quality enabled.
Based on the real-time demonstration of “Black Myth: Wukong”, we can see that the host equipped with the NVIDIA GeForce RTX 5080 (both test graphics cards have a frequency of 2797MHz) runs the game at the highest ray tracing effects, highest quality, and 4K resolution without DLSS 4 enabled, achieving a frame rate of 21fps with a power consumption of 355W. However, after enabling DLSS 4, the frame rate skyrockets to 192fps, with power consumption dropping to 298W. It is worth noting that enabling DLSS 4 also increases CPU usage.
DLSS 4 marks the most significant upgrade in the AI model field since the release of DLSS 2 in 2020. On this basis, DLSS ray reconstruction technology, DLSS super resolution, and DLAA (Deep Learning Anti-Aliasing) have all been integrated into the real-time computing Transformers model, replacing the previous Convolutional Neural Network rendering (CNN). The Transformers model has been widely adopted in commercial applications, and cutting-edge AI models such as ChatGPT, Flux, and Gemini are all built on the Transformers architecture. NVIDIA claims that the introduction of DLSS Transformers will bring higher stability, reduced artifacts, and more refined motion details, further enhancing image quality.
It should be noted that Multi Frame Generation technology is an exclusive feature of the GeForce RTX 50 series. However, the DLSS Transformers technology is not limited to use with GeForce RTX 50 series GPUs. All games compatible with DLSS ray reconstruction technology, DLSS super resolution, and DLAA technology will adopt the DLSS Transformers architecture in the future. This means that even users with older GeForce RTX GPUs will be able to enjoy better performance without incurring additional costs. Furthermore, as the new technology reduces memory requirements, the performance improvements from frame generation technology will also benefit users of the GeForce RTX 50 and GeForce RTX 40 series.
We know that after utilizing DLSS, players still need to use NVIDIA Reflex to reduce system latency for smoother gameplay. Multi Frame Generation technology can significantly enhance frame rates based on frame generation, which will inevitably increase system latency. Therefore, NVIDIA has introduced Reflex 2 technology, which for the first time employs Frame Warp technology to further reduce system latency, making player actions more responsive.
Although AI in gaming has a history of several decades, traditionally, non-player characters (NPCs) in games could only interact with players based on preset scripts. The emergence of NVIDIA ACE has revolutionized this tradition, providing NPCs with a new mode of autonomous interaction. NVIDIA ACE technology debuted in 2023, applying generative AI dialogue technology to game character development.
Thanks to the assistance of generative AI, game characters shaped by NVIDIA ACE can exhibit more natural interaction capabilities, responding in real-time to players’ text, audio, and even visual inputs, all thanks to the accompanying local small language model. The left side shows an AI NPC demonstration in “Animal Punk”, while the right side demonstrates AI teammates in the PC version of “Naraka: Bladepoint”.
In terms of efficiency for CUDA Core, RT Core, and Tensor Core, the GeForce RTX 50 series GPU based on Blackwell architecture has achieved significant improvements compared to the previous generation. Therefore, for design and creation professionals, the RTX 50 series can provide outstanding work efficiency. Additionally, this series of GPUs is equipped with GDDR7 memory, with a maximum capacity of up to 32GB, sufficient to meet the demands of high-load applications. Moreover, the RTX 50 series GPUs also feature the ninth-generation encoder (NVENC) and sixth-generation decoder (NVDEC), significantly enhancing video transcoding efficiency.
At the same time, the NVIDIA GeForce RTX 50 series supports professional-grade 4:2:2 color format capabilities and boasts efficiency far exceeding that of CPUs, undoubtedly making this series the preferred choice for video editing professionals.
In the process of incorporating neural rendering technology and popular AI computing into game experiences, three key factors must be balanced: accuracy, memory usage, and performance. In numerous application cases, NVIDIA has found that FP4 is an ideal choice, as it incurs relatively minor losses in accuracy. Furthermore, to ensure that game engine internals or the AI models running in support are more compact to minimize bandwidth requirements, even with GDDR7 memory, Blackwell employs FP4 for matrix multiplication or accumulation, thereby doubling throughput.
As seen, after adopting FP4 in Blackwell, the memory requirement for AIGC dropped from 23GB to 10GB, and the image generation efficiency significantly decreased from 15 seconds to 5 seconds.
The on-site demonstration showcased cross-platform real-time rendering collaboration based on NVIDIA NIM, FLUX ComfyUI, and Blender. Users can significantly reuse rendering materials by adding angles and objects to the scene and filling in keywords to generate the models, scenes, or images they need.
It is worth mentioning that the NIM microservice workflow designed for RTX AI PCs supports a variety of application scenarios, covering large language models (LLM), visual language models, image generation, speech recognition, retrieval-augmented generation (RAG), PDF content extraction, and computer vision. Notably, the Llama Nemotron Nano model will be optimized for RTX AI PCs and workstations as part of the NVIDIA Instant NeRF microservice, particularly excelling in directive compliance, function calling, chat interaction, coding tasks, and solving mathematical problems. The design goal of the NIM microservice is to achieve seamless integration with various AI development and agent frameworks, including but not limited to VSCode AI Toolkit, AnythingLLM, ComfyUI, Flowise AI, LangChain, Langflow, and LM Studio. Developers can conveniently download and deploy these microservices from the NVIDIA official website. Based on the NIM microservices, NVIDIA AI Blueprint provides reference implementations for complex AI workflows, aiming to assist developers in integrating multiple components, including libraries, software development kits (SDKs), and AI models, into a single application. AI Blueprint provides developers with all the resources needed to build, run, customize, and extend reference workflows, including reference applications, source code, sample data, and detailed documentation for customizing and orchestrating different components.
Additionally, NVIDIA has upgraded Broadcast, an application designed for streamers, aiming to further improve lighting effects during recording and enhance audio communication with noise reduction to meet the current streaming environment’s needs.
From the advent of the Blackwell architecture to the generative AI creation and the ecosystem built around RTX AI PCs, the GeForce RTX 50 series GPUs are dedicated to realizing NVIDIA’s grand vision in AI, gaming, and content creation. The outstanding performance of the new flagship has generated great anticipation for the launch of the RTX 50 series GPUs. In the coming time, we will conduct a more in-depth evaluation of the Blackwell architecture, including but not limited to exploring how the GeForce RTX 50 series leverages its technological advantages in ray tracing games, comparing traditional game performance, performance in creative applications, and actual energy efficiency under different scenarios. Please stay tuned.
MCers please note, due to adjustments in the public account recommendation mechanism, if you find it difficult to see articles pushed by the Microcomputer public account recently, but do not want to miss the exciting evaluation content of the microcomputer, you can move your little fingersto set Microcomputer asstarredpublic account!
