NVIDIA 发布 RTX Spark 本地 AI 平台、笔记本芯片与 DLSS 5

NVIDIA is bringing together a new Windows platform for local AI, creative work, and gaming under the name RTX Spark. The first laptops are expected to be available on October 16, 2026; compact desktop PCs are announced for November. At its core is a Blackwell graphics processor connected to a Grace CPU and up to 128 GB of shared memory. NVIDIA also promises one petaflop of AI computing performance in FP4 format, extensive CUDA compatibility, and a gaming offering with DLSS, ray tracing, and Reflex. This sounds like an unusually versatile package, but so far it is primarily a manufacturer’s proposal: prices, independent benchmarks, reliable battery life figures, and complete measurement conditions are not yet available. The product overview released today does, however, make clearer what lies behind the initially rather generous “up to” figures. NVIDIA lists two laptop configurations: one with 6,144 GPU cores and 20 Grace CPU cores, and one with 5,120 GPU cores and 18 CPU cores. For both, the company specifies a TDP range of 45 to 80 watts; depending on the configuration, the upper limits for Unified Memory are 128 or 64 GB of LPDDR5X. This is a relevant clarification, but it does not replace information on prices, specific devices, cooling solutions, or sustained performance. The advertised all-day battery life remains unquantified, as does the claim that the chip is particularly energy-efficient. Open laptop with a Windows desktop and fans visibly suggested beneath a translucent-looking base; glowing circuit patterns and the label “NVIDIA Local AI”. NVIDIA lists the ASUS ProArt P16, Dell XPS 16, HP OmniBook Ultra 16, Lenovo Pro 9n, Microsoft Surface Laptop Ultra, and MSI Prestige N16 Flip AI+ as laptop partners. The platform is also expected to appear in compact desktops from various manufacturers, including Acer, ASUS, Dell, Gigabyte, HP, Lenovo, and MSI. Pre-orders have been announced, but specific prices and complete configurations for the individual devices are not yet available. Technical details on memory bandwidth, the power consumption of individual components, and thermal limits are also insufficient to assess actual sustained performance. Press image for the NVIDIA RTX Spark platform for thin laptops and compact desktop PCs. Local AI: attractive, provided memory and software cooperate NVIDIA positions RTX Spark as a computer on which AI models and agents can run directly on the user’s own device. An agent is more than a chat window: it can break tasks into multiple steps, call programs and tools, and automate workflows. NVIDIA cites fast iteration, no ongoing API costs per request, and data that stays on the user’s own computer as advantages of local execution. For larger tasks, work can be moved to the cloud without changing the code written for it. This is a plausible usage concept, but in practice, privacy, costs, and performance depend just as much on the software, model, and specific configuration as on the hardware. According to NVIDIA, up to 128 GB of Unified Memory will support models with up to 120 billion parameters and a context window of one million tokens. Unified Memory refers to a shared memory area for the CPU and GPU. This can help with large models because multiple separate memory pools do not need to be populated; however, by itself, it says nothing about memory bandwidth or the speed of a specific inference workload. The advertised one-petaflop FP4 performance refers to a lower numerical precision that can make AI computations more efficient, but it cannot readily be equated with performance at higher precision or with an independent application test. NVIDIA names CUDA, TensorRT, and NVFP4 as development and inference tools. CUDA is the widely used NVIDIA platform through which software runs computing tasks on GPUs; TensorRT optimizes models for inference, while NVFP4 refers to an NVIDIA implementation of computations using particularly compact FP4 numerical precision. NVIDIA promises that applications developed on RTX Spark will also run unchanged on other NVIDIA GPUs. For developers, this shared software foundation is a key part of the offering; whether a project can actually run without modifications nevertheless depends on the libraries, drivers, and functions used. At launch, agent programs such as GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi, and WorkBuddy are intended to simplify setup, suggest suitable local models, and mediate between local operation and cloud services. NVIDIA Switchyard is intended as an open-source library for routing between these models; Hermes is expected to integrate it later in October, while Perplexity Computer and GitHub Copilot use their own routing solutions, according to NVIDIA. Such hybrid approaches can use local computing power for routine tasks and switch to cloud models when more capacity is needed. They do not automatically make execution free or entirely local, however. For security, NVIDIA cites OpenShell and Microsoft Execution Containers, or MXC. MXC is intended to isolate agents at the operating system level and make them controllable; according to the manufacturer, OpenShell simplifies integration and adds credential protection and security monitoring. OpenClaw and GitHub Copilot are already expected to support MXC, while Hermes Agent is expected to add the controls by the end of October. Microsoft also describes the platform as a foundation for managing agent permissions and behavior in the Windows environment. The effectiveness of these safeguards in everyday use remains to be demonstrated through specific applications and security assessments. Nemotron 3.5 Bolt 74B-A9B is due to be released on October 15. According to NVIDIA, the Mixture-of-Experts model comprises 74 billion parameters but activates nine billion of them per processing step. With 2-bit quantization and quantization-aware training, it is intended to fit in 24 GB of memory. This combination targets agents that program, research, or use tools locally; the claim that the model is faster than many widely used dense models while also being more capable than comparable MoE models comes from NVIDIA and is not supported by independent comparative tests. The named runtime environments are llama.cpp, vLLM, Ollama, Unsloth, and LM Studio. NVIDIA describes the model, including its training data, weights, and recipes, as open source. A smaller Nemotron model, Bolt E4B 4GB, is announced for October 23. It is intended to serve as the third generation of a Small Language Model for NVIDIA ACE in game characters and respond to natural conversations, tool use, and the current game state. NVIDIA highlights its low memory requirements and short response times, but comparative data is also lacking here. As earlier examples, the company cites a public test featuring AI teammates in PUBG and a voice-reactive companion in the game Aniimo. Llama Windows provides access to local models and recommendations from the Hugging Face Hub via the Windows taskbar. The Llama Windows app is intended to make models from the Hugging Face Hub discoverable and enable their local use through llama.cpp without the command line. According to NVIDIA, llama.cpp is also expanding its support for decision models, which are intended to help agents with classification and choosing the next step. However, the claimed up-to-200-times-faster decision-making compared with traditional agents is not a reliable basis for comparison without specified test tasks and measurement conditions. NVIDIA PAIR, a tool for connecting compatible computers on a local network, is expected to leave beta, support llama.cpp, and distribute inference workloads across multiple systems. The software is announced for Windows, Linux, and macOS; direct model downloads from Hugging Face and configurable inference settings are among the stated new features. Developer tools for Windows on Arm are also part of the initiative. NVIDIA and Microsoft cite a native PyTorch CUDA build, a native Python build with Python 3.15, and a Conda installer with more than 2,000 packages. Availability is already in place for 97 percent of the 15,000 most important Python wheels; NVIDIA intends to work with the community on the remaining packages. CUDA, PyTorch, and the CUDA libraries are expected to be available on RTX Spark. In addition, NVIDIA is providing so-called Skills for CUDA-X tools, intended to teach agents how to work with development libraries. The MCP interfaces for NVIDIA App, Broadcast, PAIR, Omniverse, and RTX Remix also aim to connect agents with specific programs and functions. Press image for the announced keynote by NVIDIA CEO Jensen Huang at GTC Berlin. Creative software, video, and the NVIDIA App For selected Adobe applications, NVIDIA promises up to twice the performance. Photoshop is set to gain GPU-accelerated compositing as well as improvements to Remove Tool and effects such as Light Leaks. In Premiere, the focus is on playback, rendering, color correction, and AI features such as Enhance Speech and Object Mask; Substance 3D Painter is expected to accelerate high-resolution mesh baking as well as opening, saving, and exporting large projects, among other tasks. The optimizations for Photoshop and Premiere are initially expected to be available in beta, with updates for Substance 3D to follow. Without details on the tested versions, projects, and systems, “up to 2x” remains a manufacturer claim, not a basis for a purchase recommendation. NVIDIA Broadcast 2.3 is expected to run natively on RTX Spark and add Video Super Resolution, Frame Generation, new Studio Voice profiles, and an offline file processor. Unlike real-time effects, a file can be processed afterward, which according to NVIDIA should allow certain effects to be delivered at higher quality. A new MCP server and a Skill are intended to let agents control Broadcast settings and process files. ComfyUI is getting RTX Video nodes for Frame Generation and HDR, as well as EXR export. As an example, NVIDIA cites converting a 1080p SDR clip at 30 frames per second to 60 frames per second and HDR, before exporting the result as EXR for color correction. These are workflow examples, not image-quality measurements. In Blender 5.3, DLSS Ray Reconstruction is expected to become available in November as beta support for RTX Spark. The technology is intended to improve ray-tracing results in the viewport and make interactive navigation easier. For video editing and camera workflows, NVIDIA also promotes hardware encoding and decoding for HEVC, H.264, and AV1, as well as direct processing of 4:2:2 camera footage. RTX Video is intended to sharpen footage, while Broadcast processes camera and microphone input in real time. For local image and video generation, NVIDIA emphasizes that no subscriptions, per-prompt fees, or queues are required, provided the models actually run locally on the device. Press image for creative software and applications for NVIDIA RTX Spark. The NVIDIA App brings together drivers, game and GPU settings, recording, performance indicators, and access to other applications. Game Ready Drivers are intended to optimize games on their release day, while Studio Drivers are aimed at creative software. The GPU Control Center lets users manage graphics settings and use DLSS overrides to enable newer models in supported games. Smooth Motion is intended to generate additional intermediate frames on GeForce RTX 40 and 50 GPUs. NVIDIA also offers cloud-based recommendations for graphics settings based on the GPU, CPU, and display, as well as automatic GPU tuning. No independent assessment of the performance impact or recommendations is available. Press image for the DLSS 5 3D-Guided Neural Rendering method. With ShadowPlay, users can save screenshots, gameplay footage, and the previous 30 seconds using Instant Replay. For manual recording, NVIDIA specifies up to 8K HDR at 30 frames per second or 4K HDR at 120 frames per second; RTX 40 GPUs are expected to support AV1 recording. The app also displays FPS, 1% lows, GPU and CPU utilization, as well as PC and system latency, and integrates Reflex Analyzer. Freestyle is intended to support more than 1,200 games with real-time filters, including RTX Dynamic Vibrance for adjusting color presentation and RTX HDR for adding HDR output in supported DirectX or Vulkan games. Other programs in the NVIDIA ecosystem include GeForce NOW, Project G-Assist, Broadcast, and RTX Remix. G-Assist is intended to process voice or text commands locally on the RTX GPU, change settings, and provide diagnostic data; GeForce NOW remains a complementary cloud gaming service. Press image for downloading the NVIDIA App. DLSS 5: more than upscaling, but no independent findings yet The technical focus of the gaming announcements is DLSS 5 with “3D-Guided Neural Rendering.” NVIDIA describes the technology as the final stage of the graphics pipeline: It receives the image rendered by the game and is intended to generate additional lighting and material details from it. The manufacturer cites color data, motion vectors, and other engine information, such as geometry, textures, surfaces, and lighting, as inputs. Unlike image generation based purely on text, DLSS 5 is intended to remain tied to the already-rendered scene. Engine data and artistic specifications are supposed to determine which image structures are preserved and which areas may change. The technique is intended to improve, among other things, light scattering in skin, light passing through hair and plants, contact shadows, and global illumination. NVIDIA describes the model as deterministic and temporally stable: identical inputs are supposed to produce consistent results, while motion vectors are intended to prevent flickering, image drift, and unstable details. Real-time operation at up to 4K on a single GeForce RTX 50 GPU is planned. Developers are expected to be able to select different models, control Structure Intensity for fine details such as reflections and ambient occlusion, and Tone Intensity for broader lighting and color responses, as well as use semantic or custom masks for individual objects. DLSS 5 is intended to complement, not replace, rasterization, ray tracing, and path tracing; according to NVIDIA, better source data should also improve the results. Video thumbnail related to RTX Remix modding and DLSS 5. NVIDIA points to eight years of DLSS development and lists Super Resolution, Frame Generation, Multi Frame Generation and Ray Reconstruction among the features introduced so far. According to the company, DLSS 4.5 can now generate up to 23 of every 24 displayed pixels using AI. This figure describes the combined reconstruction and image generation, not the quality or input latency of any specific game. The new DLSS 5 model is said to be ten times faster overall than the version shown in March. For the October update, NVIDIA cites around 118 million parameters, a 24 percent reduction in model size, 20 percent less 4K VRAM usage and half the GPU overhead compared with the September version. Independent tests, raw data and a complete comparison setup are unavailable. Press image of DLSS 5 neural rendering in NBA 2K27. NBA 2K27 Season 2 is set to receive the new model on October 16. NVIDIA promises up to 1.4 times higher frame rates on GeForce RTX 50 cards, RTX Spark laptops and GeForce NOW. For an RTX 5090, it cites up to 370 frames per second at 4K and up to 590 at 2560 × 1440; for the RTX 5080, RTX 5070 Ti and RTX 5070, it cites up to 410, 350 and 260 frames per second, respectively, at 1440p. According to the manufacturer, these figures are based on the Ultra preset, ray tracing and the full DLSS technology suite. Without a stated baseline, game scene, quality modes and breakdown of the individual features, it is not possible to determine what proportion DLSS 5 alone contributes. The feature is expected to be enabled with Game Ready Driver 616.64 WHQL; according to NVIDIA, it can be toggled under Video Settings in the game and switched on or off with F9 during gameplay or a replay. Video thumbnail related to the announced DLSS and gaming features for RTX Spark. NBA 2K developer Visual Concepts highlights settings that are intended to refine individual character areas without altering the players’ likenesses. For Cade Cunningham, Tyrese Haliburton and Jayson Tatum, NVIDIA cites changes to skin lighting, hair detail, contact shadows and fabric materials, among other things. The company describes similar effects on spectators and court staff. Peter Kavic, Senior Producer at Visual Concepts, also praises the image quality and control options. These are statements from a participating developer and the technology provider; the 4K comparisons mentioned in the text are no substitute for an independent image analysis. Press image of DLSS 5 in NBA 2K27. DLSS 5 is scheduled for Hogwarts Legacy starting November 5 and for Where Winds Meet toward the end of November. Developed by Avalanche Software and published under the Portkey Games label, Hogwarts Legacy is set in the 19th-century wizarding world and is said by NVIDIA to have sold more than 40 million copies. The update is expected to add Neural Rendering and update the other DLSS features; Avalanche is said to have adapted the integration to the game’s visual style. Where Winds Meet, by Everstone Studio, is a Wuxia open-world game—in other words, an action role-playing game featuring motifs from Chinese martial-arts fantasy. NVIDIA cites more than 100 million players worldwide, a Steam peak of over 250,000 concurrent players and a “Very Positive” rating based on more than 116,000 reviews, but provides no precise reference date for these reach figures. Video thumbnail related to the announced DLSS and gaming features for RTX Spark. Frame Generation, RTX Remix and the games at launch Alongside DLSS 5, a new beta preset for Frame Generation is planned for the NVIDIA App. According to NVIDIA, Frame Generation uses two AI models: one processes the game’s motion vectors and depth information, while another calculates Optical Flow, meaning movement in image content such as user interfaces or ray-tracing effects. The new Preset C replaces the second model with a lighter algorithm. NVIDIA promises higher frame rates with minor image-quality trade-offs and cites up to 35 percent higher performance for a GeForce RTX 5060. GeForce RTX 40 and 50 GPUs as well as RTX Spark laptops are expected to be supported. Activation is via the app’s beta features and the driver setting “DLSS Override – Model Presets”; NVIDIA recommends restarting if Preset C does not appear. Without specific comparisons of games, quality and frame rates, the cited percentages remain manufacturer claims. Video thumbnail related to the announced DLSS and gaming features for RTX Spark. RTX Remix is set to receive DLSS 5 and the faster Frame Generation mode on October 23. The modding platform allows older games to be updated with path tracing, modern materials and AI texture tools. A new ComfyUI view is intended to simplify the creation of PBR materials without having to set up a REST API connection through a node graph; batch processing and presets are expected to make larger projects easier. Modders will also be able to exclude objects from the influence of the Neural Rendering feature, mask areas and set model options for each object. As examples, NVIDIA shows its own development projects for Fallout 4 and Mirror’s Edge, citing changes to reflections, lighting responses, contact shadows and ambient occlusion, among other things. These project descriptions are not independent evidence of quality. At launch on October 16, several hundred games are expected to run on RTX Spark, including CONTROL Resonant, Cyberpunk 2077, Fortnite, Gears of War: E-Day, Genshin Impact, Hogwarts Legacy, Honkai: Star Rail, Marvel Rivals, NARAKA: BLADEPOINT, NBA 2K27 and World of Warcraft. Also listed are 007 First Light, Aniimo, Arknights: Endfield, BeamNG.drive, DOOM: The Dark Ages, The Elder Scrolls V: Skyrim via the SkyrimNET modding platform, NTE (Neverness to Everness), Wuthering Waves, AION 2, World of Tanks and World of Warships. The RTX features listed range from ray tracing, DLSS and Reflex to G-SYNC and RTX Remix; support varies by title, and not every feature has been announced for every game. Further dates are staggered: Among the Trolls is scheduled to launch on October 14 with DLSS Super Resolution and ray tracing; Crossout Remastered will receive a DirectX 12 update on October 22 along with DLSS 4.5, Dynamic Multi Frame Generation, DLAA and Reflex. Permafrost, Gear.Club Unlimited 3 and Pax Autocratica are listed as already available, each with different DLSS and/or ray-tracing features. According to NVIDIA, ARC Raiders has received the Frozen Trail content update, featuring the snowy Pendola Pass map, new ARC enemies, a Frigate, quests and additional progression systems. A PvE-only matchmaking test is scheduled to run from October 13 through October 20, alongside the existing PvPvE mode. For the game, NVIDIA lists DLSS Multi Frame Generation, Super Resolution, Reflex and RTXGI, a ray-tracing-based lighting technology. Call of Duty: Modern Warfare 4 is scheduled to offer campaign early access for pre-order customers on October 16, followed by the full launch with multiplayer and DMZ on October 23. NVIDIA lists DLSS 4.5 Super Resolution, Frame Generation on RTX 40 cards, Dynamic Multi Frame Generation with up to sixfold frame multiplication on RTX 50 cards, and Reflex. The manufacturer itself notes that actual launch times may differ because of time zones or possible outages. The campaign, multiplayer modes and DMZ—an extraction mode featuring looting teams and inventory carried over between matches—are also described as game information from the announcement. Some further release and update plans extend beyond launch. World of Warcraft is expected to be supported on October 16, and World of Warcraft: Forever on November 4. EA Javelin Anti-Cheat is announced for early 2027. PUBG: BATTLEGROUNDS is expected to follow in the first half of 2027; KRAFTON describes this as an intention, not a specific launch date. Arena Breakout: Infinite is planned for 2027. For Delta Force, limited tests are initially planned for a future season. Riot Games expects League of Legends and VALORANT to be available on the platform, but has not provided a specific date. Such commitments should not be equated with confirmed availability at the laptop launch. Video thumbnail related to RTX Remix modding and DLSS 5. According to NVIDIA, the company is working on anti-cheat support for multiplayer games so that RTX Spark users can compete alongside players on other platforms. The systems named are BattlEye, EA Javelin Anti-Cheat, Easy Anti-Cheat, RICOCHET, Vanguard and Tencent Anti-Cheat Expert. The announcement lists ARC Raiders, Arknights: Endfield, Counter-Strike 2, Dota 2, Fortnite, Marvel Rivals, NARAKA: BLADEPOINT, Rocket League, RUST, THE FINALS, War Thunder, World of Tanks, World of Warships and Wuthering Waves, among others, as supported titles at launch. However, the general announcement does not establish that every anti-cheat solution has already been fully implemented for every game. Press image of the DLSS weekly overview featuring Call of Duty: Modern Warfare 4, new games and a faster DLSS 5 model. The games overview also lists numerous projects from manufacturers and partners. BeamNG.drive is expected to be able to use the real-time physics of its vehicles on RTX Spark; War Thunder lists several thousand vehicles, aircraft and ships and promotes ray tracing, Reflex and DLSS 4.5. Arknights: Endfield combines real-time combat with the construction of supply chains and bases; DOOM: The Dark Ages and the Revelations expansion are expected to support DLSS Multi Frame Generation and Ray Reconstruction. 007 First Light, Genshin Impact, Honkai: Star Rail, Zenless Zone Zero, Aniimo, NTE and AION 2 also appear in the compatibility overview. Where the announcement cites high sales or player numbers, for example for ARC Raiders, Delta Force, NARAKA: BLADEPOINT or Hogwarts Legacy, the reference period and independent data-collection method are often missing; these figures should therefore be treated as claims made in the respective announcement. What remains unclear before purchase For creative professionals, NVIDIA lists 4:2:2 camera formats as well as hardware encoding and decoding for HEVC, H.264 and AV1; for gaming, it lists HDMI outputs of up to 4K at 480 Hz or 8K at 120 Hz with Display Stream Compression, VRR and HDR under HDMI 2.1b, as well as DisplayPort 2.1b. These ports and capabilities may vary by laptop model and need to be checked for the specific device. TDP and memory configuration figures are also no guarantee that all peak values will be achieved simultaneously or sustained. The key question, therefore, is not whether RTX Spark can offer a long feature list, but how the systems perform in independent tests. Prices, battery tests, noise and temperature measurements, real-world gaming benchmarks and comparisons of local AI tasks with similarly priced alternatives are currently missing for a purchase decision. DLSS 5 image quality and latency should also be assessed, rather than equating high generated frame rates with correspondingly high native rendering performance. Frame Generation can increase the displayed frame rate, but does not automatically eliminate latency or the requirements for the actual rendered base frame rate. RTX Spark is therefore an ambitious combination of Windows laptop, developer platform and local AI system. The CUDA foundation, high Unified Memory limits and planned software integrations could make the devices interesting for specific workflows, while compact designs and gaming features may appeal to additional user groups. Whether this makes for a good product, however, will depend on sustained real-world performance, battery life, software quality and price. Until these aspects have been measured, the “all-in-one” promise remains primarily a manufacturer announcement, not the result of a laboratory test. Still, we would be happy to be pleasantly surprised.